I'll be starting the session with the first lecture on diagnosis and staging of pancreatic cancer, state of the art, and I'm thankful to RSNA for the opportunity. So in this lecture I will discuss the imaging technique, imaging manifestations, and the current staging guidelines in pancreatic cancer, and briefly touch upon early detection and risk assessment. So pancreatic cancer, or ductal adenocarcinoma, is the fourth most common cause of cancer mortality in the United States, and arises from the exocrine pancreatic ductal system. According to a new report, the incidence of this cancer has increased since 2001, and in 2022 alone, nearly 50,000 deaths are estimated in the United States due to this disease. Now this tumor has an aggressive biology, and patients often present late due to complex regional anatomy and lack of reliable tumor markers. Surgical resection offers curative treatment option, however, less than 25% of patients have resectable disease at the time of presentation. Compared to other GI cancers, this malignancy is a poor prognosis with a five year survival rate of five to 10%. Now imaging plays an important role in the diagnosis and staging of pancreatic cancer, and different imaging techniques are available. To identify the imaging practice pattern in pancreatic cancer, we performed a multi-institutional survey of sub-specialty abdominal radiologists within the US. What we found was that nearly 93% of surveyed radiologists perform multi-phasic pancreatic protocol CT as the first line imaging investigation for diagnosis and staging of pancreatic cancer with nearly 20% of them using dual energy or multi-energy CT. On our multi-phasic pancreatic protocol CT involves acquisition of the pancreatic phase followed by the portal venous phase and an optional delayed phase. Thin section imaging, use of neutral oral contrast and multi-planar reformats are integral to this protocol and allow optimal detection of the tumor as well as its relationship to the arterial and portal mesenteric vasculature. Now the pancreatic phase is acquired slightly later than the early arterial phase and provides the most optimal contrast between the hyperdense pancreatic cancer against the background of enhancing pancreatic parenchyma while allowing for delineation of the surrounding arterial vasculature. The portal venous phase is required for detection of hyperdense liver metastasis as well as for evaluation of the portal and mesenteric veins. As we found in our survey, there is increasing utilization of dual energy or multi-energy CT and this technology allows for generation of numerous different types of image data sets. From a practical standpoint though, the images which are relevant in clinical practice include monoenergetic and the iodine-specific images. Now in our practice, we mostly use dual energy CT for pancreatic cancer patients. So the typical image data sets, 10 to PAX, from our pancreatic protocol dual energy CT have been optimized so that the least number of images are generated to provide the maximal information. So in addition to the images resembling single energy CT images, we generate low KEV monoenergetic images at 55 KEV for pancreatic phase and 65 KEV for portal venous phase. Iodine and virtual non-contrast CT images are generated for both the phases. Now coming to the imaging manifestations on CT, pancreatic cancer are hypoattenuating tumors. They are infiltrative, cause obstruction of the pancreatic duct and involve adjacent vasculature as well as the surrounding structures. A classic imaging appearance as seen on this MRCP image is the double duct sign with the obstruction of the both pancreatic and the biliary ducts. Up to 10 to 15% of pancreatic cancers are isoattenuating and hard to identify on CT. Now these tumors can cause duct obstruction and lead to what is called as a duct cutoff sign. This duct cutoff sign, as you can see in multiple images here, is when early pancreatic cancer causes obstruction of the pancreatic duct and upstream dilatation. When seen on imaging, this is an important sign as it is ominous and the site of transition where the duct cutoff is seen needs to be aggressively evaluated with EOS or MRI. Now this is because identifying pancreatic cancer at this stage provides the best opportunity for curative surgical resection and improved prognosis. Now multi-detective CT has high diagnostic performance for staging, particularly for evaluating vascular invasion. However, CT does have limitation such as identification of isoattenuating tumors, lymph node characterization, detection of small liver metastasis, and recognition of peritoneal disease. I want to briefly go over the recent technology, dual-energy CT, how it allows improved diagnosis of pancreatic cancer. Multiple studies have shown that the low-energy, mono-energetic images in the 45 to 55 keV range improve detection of pancreatic cancer during the pancreatic phase by providing higher pancreas to tumor contrast. Iodine images, both gray scale as well as colored maps, improve visualization of the primary tumor and provide higher radiologist confidence. In presence of patients of metallic or biliary stents, those who have undergone endoscopic intervention, iodine images mitigate the stent-related artifacts, thereby enhancing visualization of tumors in the presence of metallic or plastic biliary stents. Improved delineation of the tumor and enhanced visualization of the vessels is important for better treatment planning by improved margin delineation. Also there are recent studies which have shown that iodine images improve identification of small enhancing peritoneal nodules and improve diagnosis of peritoneal carcinomatosis in stage four pancreatic cancers. So to summarize, dual-energy CT, a combination of low-keV, mono-energetic, and iodine images in pancreatic cancer improves tumor conspicuity, allow better delineation of margins, allow enhanced visualization of the arteries and veins, improve detection of tumor in presence of stents, thereby providing improved tumor staging. So what's new with imaging technique? I wanna go over two concepts here related to phases of CT acquisition. I'm gonna start with one phase versus two phase. So in two recent research purpose, we found that the low-keV image generated from a portal venous phase acquisition could replace the traditional pancreatic phase and provide comparable diagnostic accuracy for detection of both solid and cystic pancreatic lesions while allowing for significant radiation dose reduction. Now this single-phase twin pancreatic protocol CT with reduced dose is desirable in patients undergoing surveillance such as those with pancreatic cysts or those at high risk for developing pancreatic cancer. Let's look at a different perspective of increase in number of phases. In a recent paper in AJR, Fukukura and colleagues compared triple-phase CT with traditional dual-phase CT for detection of small PDAC. In a multi-reader study involving a cohort of 97 patients, they found that addition of a delayed phase to the pancreatic and the portal venous phase increased sensitivity for small tumor detection without loss of specificity, especially for isoattenuating tumors partly related to the delayed hyperattenuations of these lesions on the delayed phase. As you can see in the example below, there's a hyperattenuating tumor seen in the delayed phase which is isoattenuating to the pancreatic parenchyma on the pancreatic and the portal venous phase. Now coming from the diagnostic perspective on a briefly discussed early detection and risk assessment. Starting with the early detection in this interesting paper, Chari and colleagues looked at CT scans done in patients three to 36 months before diagnosis in pancreatic cancer. And they found that nearly 12 to 18 months before diagnosis, progressive and increasingly frequent changes occur on CT scans and standard CT was 86% sensitive for detecting tumor three to six months before diagnosis. Now this emphasizes the importance of early detection of pancreatic cancer which is critical as this has significant survival benefits. Now I look forward to a talk by Mike Rosenthal on AI and radiomics to address this important issue. The second concept is risk assessment. Now there has been a lot of interest lately in exploring body composition metrics made available by CT to help predict risk and outcomes. One such parameter is sarcopenia which refers to reduction in skeletal muscle quantity and quality. Now there's extensive research on this subject and there is increasing evidence that pancreas sarcopenia is considered a poor prognostic indicator in cancer patients including in patients with pancreatic cancer. Now here are two clinical examples we encountered in our practice with pancreatic cancer and you can see the overall survival in a patient with more muscle bars on the left when measured at L3 vertebral level was significantly higher compared to the patient with sarcopenia who had a poor survival. Now quickly going over the staging perspective, there are two major systems for staging in pancreatic cancer, the American Joint Committee on Cancer staging system and the clinical or radiological staging system based on NCC and consensus report. In this clinical, in the interest of time I want to focus on the clinical staging system. In this staging system pancreatic cancers are categorized radiologically on a continuum from resectable, borderline resectable and locally advanced or undesectable tumors according to the involvement of adjacent structures and the presence of distant metastasis. This categorization is crucial for determining treatment strategies as resectable tumors often undergo upfront surgical resection while tumors with distant metastasis receive systemic chemotherapy. I mean patients with borderline resectable or locally advanced tumors, preoperative neoadjuvant chemotherapy is often administered to render the tumor resectable. Now determining the extent of disease on imaging is important as the goal of surgery in pancreatic cancer is R0 resection. Margin positive resection strongly predicts early recurrence and shorter survival. Now in addition, despite the advances in surgical techniques, postoperative mortality can be high, particularly when RTL reconstruction is undertaken. Therefore the NCCN guidelines recommend standard terminology for vascular involvement as it has huge impact on the susceptibility. So let's briefly go over those terminologies. Starting with RTL involvement, a less than 180 degrees of tumor contact with vessel circumference is considered as abutment. More than 180 degrees of tumor contact with vessel circumference is called encasement and this imaging sign that is encasement predicts vascular invasion with a high degree of specificity. Additional findings suggestive of vessel invasion include tumor causing deformity, narrowing, irregularity, and thrombosis. Similar terminology can be applied for venous involvement as well. Going over the different categories, resectable tumors are defined as those tumors without distant metastasis and lack of RTL or venous tumor contact. Venous abutment without contour irregularity also makes the tumor upfront resectable. The next category of borderline resectable pancreatic cancer refers to tumors with greater vascular involvement, that is encasement of veins which are reconstructable and abutment of arteries such as the supramucentric and the celiac common hepatic artery. Locally advanced tumors have even greater degree of vascular involvement such as encasement of SMA and celiac axis as it's seen in this example or even more extensive involvement of the veins. Distant metastasis in pancreatic cancer occurs to liver, lymph nodes, and peritoneum as seen in these examples. Now the clinical staging for resectability, that is differentiation into these different categories is performed in the multidisciplinary tumor board and we don't put this in our reports and this requires an involvement of the radiologist, medical, surgical, and the radiation oncologist. Imaging description is of utmost importance in patients with pancreatic cancer. So we have Dr. Olga Brook who's gonna speak on structured reporting templates in patients with pancreatic cancer. So with that I would like to summarize what we just learned in the last 11 to 12 minutes. Pancreatic protocol CT is the preferred imaging technique in the diagnosis of pancreatic cancer with increasing utilization of dual energy CT. Pancreatic duct cutoff is an ominous sign. Early detection of this tumor is critical for improving survival and knowledge about the susceptibility criteria and imaging descriptors is crucial. Thank you for your kind attention. I wanna share with you all some pros and pitfalls in pancreatic cancer response assessment and I wanna do that through a series of three case discussions. So this is case one. We have two different patients with locally advanced pancreatic cancer undergoing neoadjuvant therapy as well as restaging scans and we see patient A here and draw your attention to the unscented process where the tumor is pre-neoadjuvant therapy and post-neoadjuvant therapy. And here's patient B. These are the coronal reformatted images and again pre and post. The question is which patient should undergo surgery? In other words, which patient do you think most likely responded to neoadjuvant therapy and most likely will have a R0 or margin negative resection? Is it patient A, patient B, neither or both patients? Give you just a couple of seconds to take a look at the images. I think most of you would say that it's A, patient A and indeed if we look at that on the pre-neoadjuvant therapy CT, the baseline CT that we see tumor encasing the vessel, the SMA and on the post-neoadjuvant therapy, the tumor size has decreased and there is also reduced tumor contact with the vessel. And that indicates response. In contrast for patient B, there is actually increasing size of the tumor as well as increased vascular contact. And indeed patient A underwent R0 resection with excellent histological response. And this really illustrates the, probably most important pitfall in response assessment in these patients is that it's very difficult to differentiate fibrosis from viable tumor following neoadjuvant. And this study as well as some other studies have shown that the CT accuracy for predicting resectability decreases following neoadjuvant and it's much lower percentage, 58% compared to the control group which is those who undergo upfront resection. And this is really related to an overestimation of vascular involvement and underestimation of histological response. Fortunately we do have some imaging predictors of response following neoadjuvant. And this study showed that partial regression of tumor contact with any peripancreatic vessel, SMV, celiac, so on, predicts R0 resection. And if you see a change in tumor size and a change in post-treatment CA99 level predicts histological response. So in that index case, if you remember, even though there's persistent soft tissue contact with the vessel, that the fact that it has decreasing size and decreasing vascular contact predicted R0 resection. A related companion case is shown here on the pre-neoadjuvant treatment CT. You can see the tumor encasing the SMA. On the post-neoadjuvant therapy CT, the tumor size has decreased, but more importantly, the solid tumor has been replaced by this hazy attenuation. And this patient also underwent surgery with a R0 resection and excellent histological response to treatment. The haziness may be residual tumor, fibrosis, or inflammation. And this study showed that the accuracy in the prediction of vascular involvement following neoadjuvant therapy was improved by not considering this perivascular haziness as a sign of vessel involvement. What are some of the other predictors for R0 resection? What about if you just look at the post-neoadjuvant treatment CT? This study showed that if you see tumor contact of vessel that's 180 degree or less, without vessel contour deformity and or if the length of the tumor contact is less than two centimeter, that is a good sign for predicting R0 resection. And here it just illustrates this concept where this tumor is abutting the portal SMV confluence, it's not deforming the vessel, and it's a relative short contact with the vessel. And this patient also had R0 resection with excellent histological response. Now in addition to these CT imaging findings, FDG PET can also be very helpful in differentiating between fibrosis and a viable tumor and in predicting pathological response as shown here. You see a tumor in the pancreatic head, and in this case, pre- and post-treatment PET showing near complete metabolic response. And indeed, this recent paper that looked at over 200 patients with borderline resectable or locally advanced pancreatic cancer who underwent neoadjuvant therapy, that the metabolic response on PET was the strongest independent predictor of pathological response. And while FDG PET is really not standard practice, that this paper does suggest that it should be considered for patients undergoing neoadjuvant therapy. Now let's move on to case two. There are three different patients with pancreatic cancer undergoing neoadjuvant therapy. And during the therapy, they had interval staging study and showed a new liver lesion in each of these three patients. So the question is, which of the below is most suspicious for liver metastasis? I'm gonna just give you a few seconds to have a look at the images. The answer is A, this patient had metastasis and of course is no longer a surgical candidate. And the other two cases were abscesses which resolved to following treatment. And differentiating between liver abscesses and metastasis is another common pitfall. Liver is a common site of metastasis in pancreatic cancer. And abscess is also quite common, especially after biliary intervention because of this disruption of the sphincter of Odi, predisposing the GI tract bacterial entering the biliary tree. Now there are some imaging signs that can help differentiate between the two. And this is the double duct sign which is really referring to, that that's a sign for abscess. It's really referring to a hypo-enhancing central area surrounded by first a hyper-enhancing ring representing inflammation, followed by this hyper-hypo-enhancing zone representing edema as shown in this case. And you can see it perhaps a little bit more clearly on MRI. And again, recognizing this zone of edema, that's a really useful sign that points to abscess. Another sign is the so-called cluster sign which is really referring to these multiple small lesions that kind of aggregate and coalesce into a single large abscess cavity, which is, this was one of the image on the set of three images. This is an abscess. And again, here you can see the coalescing of these small lesions into a large abscess cavity. And if you look at metastasis and abscess side-by-side, in metastasis you see a ring of enhancement. And sometimes you can see this perilesional enhancement, and that really points to metastasis in comparison to abscess. In this double duct sign, again, you see this zone of edema at the outermost layer. Now let's move on to the last case, which is a 57-year-old man status post-WIPO for pancreatic surgery and, or pancreatic cancer undergoing surveillance study. And here's a baseline study following WIPO. And this is a new six-month follow-up surveillance study. The question is, what do you tell your colleague at Tumor Board, is it A, that you see stable findings between the two studies and just continuous surveillance, B, there's increasing pancreatic ductal dilatation, but likely related to benign stricture, or C, that there's increased ductal dilatation that's suspicious for recurrence, and you would suggest something else, like a FDG PET scan to further evaluate? In this case, the answer is C, and indeed, this patient underwent a FDG PET, and what you see here is a focus of hypermetabolic focus at the pancreatic jejunostomy anastomotic site, that's where the recurrent tumor is. Now FDG PET is very useful in equivocal cases where you have CT findings that are equivocal, or if you do not see any CT findings, but there's elevated CA19I, it is helpful in differentiating between tumor from fibrosis or benign stricture, and PET does detect a recurrence earlier than CT, 96% for PET versus only 57% for CT. It's especially useful to detect a recurrence at the anastomosis, as you know that the altered anatomy can both mimic cancer as well as obscure a tumor at the anastomosis. Now a related pitfall post-surgery is, as you know, that soft tissue is commonly seen surrounding vessels such as the SMA and SMV, and prominent lymph nodes are also very common, and in these cases, it's really hard to tell surveillance is needed to establish stability. So here's an example, you see a small amount of soft tissue between the aorta and SMA, and this patient was surveillanced for many years, and this was stable, indicating that this is just post-surgical changes. In contrast, this patient also had a small amount of soft tissue anterior to the vessel on the baseline scan at a six-month follow-up. It definitely has increased, and this is recurrence, proven recurrence. So just to summarize the take-home messages, in the neoadjuvant setting, any decrease in tumor size or vascular contact suggests a histological response, so surgical exploration should really be considered. FDG-PET may improve that response assessment. Beware of liver abscesses, mimicking metastases, imaging signs we discussed, double target sign and cluster sign that point to the diagnosis of abscesses. And finally, following surgery, beware of altered anatomy and post-surgical changes which can both mimic or obscure recurrent tumor. And again, FDG-PET could be helpful in equivocal cases. And with that, I wanna thank you very much for your attention. So I'm gonna talk about structure reporting. I don't have to tell this audience that staging pancreatic CTA is a great study. It allows superb delineation of vascular anatomy. It's relatively accurate for local tumor extension. And most importantly, as was shown previously, it determines eligibility of patients for tumor resection and thus determines patient's outcomes. Extremely important. And this management is really determined by those characteristics, post-tumor, and vascular characteristics of the patient specifically. Now, we need to report those, right? So how would surgeons know? A lot of my colleagues used to say, I mention all the relevant details anyway with non-structure reporting. I don't need your structure reporting. It's just too tedious, too hard, don't need you. Well, we tested that hypothesis a while ago now. And we found that no, that's not true. With non-structure reporting, we're not discussing, we're not mentioning a lot of features that needs to be mentioned. And yes, those are personal negatives, but they still need to be reported. And with structure reporting, there is definitely improvement in all of those. Now, the question that comes through is do we actually need that level of detail when we're reporting about those patients? Do we need to report all those personal negatives? So what we did, we asked surgeons. And this is just one of the questions that we ask, and most important question, though. Can you determine tumor susceptibility based on the report? And here you have experienced surgeon with more than 20 years of experience that just read the report. And on non-structure reports, he could determine tumor susceptibility in 42% of the cases. And the same studies, but done via structure reporting, he could do it in 60% of the cases. So definitely there is need for structure reporting, even for the most experienced surgeons that will still look at the images, but the structure reports are helpful to them. So overall, disease-specific structure reports, they increase in number of reported descriptors. They're easier to extract needed information. They facilitate ability of the surgeon to decide on tumor susceptibility. So yes, the answer is yes, we do need that level of detail. And so a few examples of what we would not routinely report, at least I would not routinely report in general abdomen and pelvis. Replaced tritepidecartery, it's a normal finding, present in 20% of patient population in general. But if it's present in patient that needs Whipple, it makes Whipple technically very challenging, prolongs our time by 60 minutes, and if it's not managed or detected preoperatively, it can be a disaster during the surgery. So it certainly needs to be addressed ahead of time. Additional vascular findings that are worth mentioning, here is just a few. Celiac stenosis, right, we see it all the time, but for patients that will undergo Whipple, it's very important to mention that this is present, it needs to be addressed. Median arcuate ligament syndrome, similarly, SMA stenosis, again, if you don't mention this, if they don't treat this ahead of time, they don't address this, that may cause anastomotic leak, because the anastomosis is not gonna heal appropriately. So again, those things need to be mentioned ahead of time. So this is surgical literature, right, and yes, it's a little old, but not that old. And it states that visceral angiography, right, angiography should be considered routine before Whipple, particularly in surgical residency training. This is a paper from 1998. Detection of vascular anomalies, CTA is quoted at 67%, pretty horrible, I would say, and DSA at 100%. And I would say those numbers are, we know as radiologists that those numbers are wrong, and CTA can be as accurate as angiography for vascular depiction. It's really just a matter of reporting, structure reporting in this case. So I was very happy to see this tweet by my favorite abdominal radiologist, Twitter, saying that she hated structure reporting when she started, but now she thinks it's stupid not to use them to make sure you don't miss anything, and really, are you really that stupid not to use them when you can just get some help? So this is really some help that you can get. And so, pancreatic cancer DFP of SAR came out with this consensus statement together with American Pancreatic Association in 2014 that they produced this template that we can all use. And here is a depiction of arterial involvement, venous involvement, and distant disease. All great. This was then followed and incorporated into NCCA guideline in 2017, and basically, the guideline states that use of uniform reporting ensures complete assessment and reporting of all imaging criteria essential for optimal staging, and thus, will aid in determining optimal management. And so basically, NCCA now routinely recommends to use this template, exactly that template that I showed you before. Now, what's the problem? The problem is 37 fields, right? It's a little long, right? And together with how many series you had, Avinash? That's not surprisingly, those studies stay on the list for a little while. And just to remind you, the study that we, where we show the benefit of structured reporting for pancreatic cancer, we only had 12 fields, right? So from 12 to 37, that's a little bit of a jump. And this is something that is known in the literature as a feature creep, right? So we're starting with user needs. We ask the surgeon what they actually need to determine receptability, we got this. But then, we invited stakeholders, and everybody has ideas, right? I want this, and I want this, and I want this, and then we become, is this really a monster, which is kind of hard to use. And now, Dr. Kambodakone did a survey under, of pancreatic cancer, DFP of SAR, surveying 35 major academic institutions, and all radiologists, I believe, if I'll ask all of you, will also say you believe in benefits of structured reporting. However, how many of you actually using it? Well, in this survey, it was 61%, and that's in the academic, among academic radiologists who report pancreatic cancer. And why? Because it's complex, because it interferes with efficient workflow, right? So nothing new here. So you have this tug of war between the efficiency, right? We need to read all those studies, and the quality is the evidence. But again, the evidence is really not for the 37 one, it's really for the 12-item template. And so what is the solution? Again, thanks to pancreatic cancer, DFP, we have the leaders here, that came up recently, just a month ago, specific template for specific situation, right? So post-op patient does not necessarily need all this huge, full reporting, right? So this is a specific template for post-op patients that only has eight items, much shorter, but still provides all the necessary diagnosis. So I would refer you guys to abdominal radiology, last month, it just came out, fresh. And that can be useful, and that would facilitate the efficiency of reporting, but still reports all the important stuff that surgeons needs to know, and oncologists in those cases. So in summary, pancreatic cancer staging CTA requires structured reporting, absolutely, for pertinent positives and negatives. The full template is endorsed by NCCN, so I would highly recommend that you use it for your surgeons, for oncologists. But do use limited template for post-op and inoperable patients. Thank you. Thank you so much, Avinash, and thank you all for the opportunity to speak today. I appreciate it. So today I'm gonna talk about a very, very exciting year in AI and pancreatic cancer, because there have been so many interesting studies that have come out this year, and I'm really excited to share several of them. I'm gonna talk about four areas where I think that AI is really going to affect the life of the radiologist who is reading pancreatic cancer studies in particular. The first is computer-aided section and classification. Having AI systems that help us find small lesions that we might miss while we're looking for an appendicitis or gallbladder that's gone awry. But when we get a little red flag there, we say, oh, yeah, that's something we actually need to get an MRI or an endoscopic ultrasound done. Risk stratification, all of those cystic lesions that haunt our impressions. There are some new AI tools that are showing very promising results for risk stratifying those. Second group is new risk biomarkers. Organ radiomics has come up in two major papers this year. Very exciting, shows some suggestions in early detection. Muscle wasting is my personal area of greatest excitement. I'm gonna talk about that. And then wrapping that all into some opportunistic screening activities. The third is new high-risk screening groups. So if you do high-risk screening for familial pancreatic cancer or genetic-risk pancreatic cancer, now you know that's a relatively small pool of people. There are some new risk factors coming online, driven in part by AI tools that may be able to identify new groups of people who could go into those screening programs and benefit based on enriched risk pools. And finally, treatment allocation. Trying to understand if we can do a better job with those vascular margins. There's been some progress on that this year. And metastatic risk, although I'm not gonna have time to talk about metastatic risk today. So computer-aided detection. This has been around since the age of mammography, back in the late 1990s. And the idea is having an AI system look at a study that you're reading and flag at-risk lesions. This particular study came out in Radiology this year, fairly recently, by Chen et al. It was a study of AI detection of pancreatic cancers from clinical CT scans. Retrospective case control design using the Taiwan National Database, which captures all of the occupants and citizens of Taiwan. They thoughtfully used separate training and testing sets. The sample sizes were pretty good. And they used an ensemble of five different convolutional neural networks for classification, which means they trained basically five different random models that use a neural network that works kind of like the human visual system to say, is this going to be a cancer? Is this not going to be a cancer? And they found remarkable performance. This ROC curve shows that each of the networks individually performed very well, but the ensemble taken all together had an area under the curve of about .95, corresponding to about 91% accuracy with 90% sensitivity and 93% specificity. And the performance was not significantly different than the radiologists who were looking at the same studies. Now there wasn't a formal test of equivalence there, but it certainly speaks to the quality of the systems that are coming out at this point. The next study that I'm gonna talk about is risk stratification for cystic lesions. This was a single institutional retrospective study of two 14 subjects out of Hopkins, led in part by Linda Chu, Elliott Fishman and crew. The cases included 64 IPMNs, some MCNs, seroserous cystadenomas, SPENs, and cystic neuroendocrine tumors. They did manual segmentations of the pancreas and of the lesions, and then used radiomic analysis of the pancreas and of the cystic lesions. Radiomic analysis generates a huge quantitative feature vector list that looks at things like how dense is it, how solid is it, does it have stripes, does it have bands, does it have dots in it, and does that in a very comprehensive way. They then used something called random forests, which basically try to take the original set and find the feature that best divides it into two risk pools, and then in each of those risk pools finds the next best feature, and keeps marching down the parameter list until it has only one patient in each bucket, and that's a random forest. Takes a whole group of them, and you have not just random trees, but random forests. They found that this AI method actually outperformed radiologists, producing an area under the curve of .94 versus .895 for the radiologists. Again, really excellent, and offers us some hope that perhaps we can do something with these pancreatic cystic lesions other than just annual follow-up for forever. And I do recognize that there's nuance there. New risk biomarkers. So this has been a truly fantastic year for radiomics in pancreatic cancer. Up until this time, it was viewed as a fairly unreliable tool, and these two new papers have really given a fresh hope to this. The first of the papers came from the group at Cedars-Sinai and Kaiser Permanente, Southern California, led by Cresci et al. as first author. This was a retrospective case control design. They looked for cases for CT scans that were acquired six to 36 months before the clinical diagnosis, and they also got the paired at-diagnosis CT scans, and they also acquired healthy controls that did not have any pancreatic abnormalities. It was a relatively small sample size, training and validation in 22 cases and 22 controls, external testing in 14 cases and 14 controls. So a small sample, nonetheless, they were able to get through the manual segmentation, did radiomic extraction of a very large number of features, and came up with a binary decision model using some fancy techniques, and arrived at an 86% classification accuracy. And you can see the confusion matrix here shows that it actually did a pretty good job of binning healthy people into the healthy bin and cancer people into the cancer bin. For such a small study to arrive at significance, it's a positive sign, and that it spans the six-month to three-year window means that there's some hope that that's actually a signal that could matter in terms of outcomes for our patients. I'd love to have an extra six months for all of our patients to get an earlier stage. The second study that came out was from the Mayo group with our colleague, Mukherjee, and of course, Ajit Gunkha, and many other people from that group. This was an exciting study that used a similar kind of design retrospective. They used it three months to 36 months before diagnosis, but with a median of 398 days of the pre-diagnostic scan to the diagnosis. Training and validation in 110 cases, 182 controls. Internal testing on 45 and 83 controls. Excuse me. About 40% of the cases had subtle pancreatic abnormalities on radiologist review that had not been recognized at the time. This was, again, all done with manual segmentation, and they used a smaller feature set of 88 radiologic features. They compared four different commonly used classifiers that are often in the machine learning domain, and I'm not gonna go through the details there, but they arrived at an area under the curve of .98 for the best performing models versus .66 for the human readers. And you can see this area under the curve, again, is really, really promising. This technique produces these feature maps here where you can see kind of a heat map on the pancreas showing these values at different voxels, and when we put these all together, we get a really extraordinary classification model. Now, we still need to know how all of these techniques perform in a real-world prospective cohort because these are all enriched to about one-to-one cases and controls, and if we were seeing half of our cases have pancreatic cancer, that would be a pretty bad time. And so we've gotta really show that we can live in the 20 per 100,000 domain and not have to do an extra 10,000 MRIs on people who have benign findings before we're ready to adopt this. But both of these papers give very strong evidence that pancreas radiomics may actually be a path to early detection for us. Another study that I'm gonna share is near and dear to my heart, and this data is under review right now at a journal, so full disclosure there. This is a retrospective matched case control design that we did from Dana-Farber in conjunction with our friends at Kaiser Permanente, including Bedi Khan and crew. We have 904 pancreatic cancer cases and about 2,400 matched controls, and we looked at CT scans collected six to 60 months before diagnosis here, a little bit farther in the past than the radiomics studies. We looked at body composition analysis as measured at the L3 level using a fully automated system, so a system that accepts complete abdominal CTs, finds the L3 level, processes them, spits out numbers, and so you could run this on your PACS on absolutely every study that you did without having to touch it. This was a paper that I did with my former resident and now colleague from Duke, Kirti Magudia, who perhaps is here. We used demographic normalized percentiles, z-scores, to look at how these people changed over time relative to their peers, and you can see in the pancreatic cases, diagnosis is over here on the right at zero months. Beginning at 60 months beforehand, there's a continuous decline in skeletal muscle percentile across this entire interval. It becomes statistically significant at 18 months, but we don't really know what happened before this, and so we just know that there is a significant trend across this entire interval, whereas matched controls stayed flat. This supports the prior work that Suresh Chari and crew had done at Mayo, suggesting that muscle wasting is probably another leading indicator of pancreatic cancer, and also goes along with data found in the blood in the nurse's health study, my colleague Brian Wolpin had found several years ago. Again, really exciting evidence that perhaps we have two imaging-based biomarkers that could offer a window to early detection. So there's no effective early screening approach in the general population right now. We use the CAHPS protocol for high-risk individuals, which includes annual MRI and endoscopic ultrasound, but we really need to do better. Opportunistic screening may be a pathway for this, so Perry Pickhart and Ron Summers have done an extraordinary, extraordinary job of leading the way in opportunistic screening through fully automated analysis of CT scans. I really love seeing their papers. I love being able to occasionally review their papers, and I think that pancreatic cancer needs to do this. The tools we need to do that include automated multi-organ segmentation. This is a tool that we've developed to automatically segment the pancreas, but once you can do that, you can now do, for example, automated radiomic analysis of 100,000 cases, and that's a fine thing to do, and we can start talking about how screening looks like in real unselected populations. Similarly, body composition analysis. I think this is really ready to go now and can be deployed on clinical systems very easily. We just need to accept and approve our driving indications, and I think that cardiovascular disease and cancer look like the two that are gonna win out, but these two combined together could mean opportunistic screening from studies that we're doing otherwise. New screening groups, also very exciting. I'm gonna treat this fairly lightly and just say that people are doing work on extracting features from clinical factors like new-onset diabetes, including the lovely work by Suresh Chhari's group again, up in Mameo in gastroenterology, showing that the NPAC score can improve detection in patients with new-onset diabetes, and another paper from one of my collaborators, Chris Sanders, that's under review right now, showing that ICD code trajectories in the EMR are also predictive of a future diagnosis of pancreatic cancer. If we can find these at-risk groups in the general population and enrich them to the same levels of the high-risk families, so getting up to about 1% incidence per year, then it would be a pretty strong rationale to pursue the same kind of MRI endoscopic ultrasound approach for screening these people. Wrapping up here, treatment allocation, operative risk, this was a lovely paper that came out at the end of last year in AJR, a single institutional study of 153 resected PDAC cases looking at operative outcomes, and basically asking the question, who should be going to the OR? They looked at 153 cases, manual segmentation, 489 radiomic features, good division of cases for training and validation. What they found was that if you added radiomics features of the tumor and vascular margins to clinical factors, you actually improved the predictive capacity of these parameters from 0.68 area under the curve to 0.74, again suggesting that maybe we can do a little better than the NCZEN guidelines alone. So in conclusion, there are new AI-based tools that are likely to be available to flag potential lesions, identify some at-risk phenotypes, and assist with some of these perioperative assessments that we work so hard on. AI may be able to improve our ability to identify at-risk individuals and target them for screening within the next three to five years. Optotistic screening with body composition and organ radiomics looks really promising after this year, and I'm hoping that in the next couple of years we're gonna see papers coming out prospectively, and I think radiologists are going to be very important gatekeepers and stewards of these technologies. Thank you all so much for your time. There's a very large team that have worked on all of these things that I've shown you. I want to give them credit. Thank you.

pancreatic cancer

diagnosis

imaging

multi-phasic CT

structured reporting

artificial intelligence

radiomics

treatment planning