Automatic Personalized Impression Generation for PET Reports Using Large Language Models

Xin Tie, Muheon Shin, Ali Pirasteh, Nevein Ibrahim, Zachary Huemann, Sharon M. Castellino, Kara M. Kelly, John Garrett, Junjie Hu, Steve Y. Cho, Tyler J. Bradshaw

In this study, we aimed to determine if fine-tuned large language models (LLMs) can generate accurate, personalized impressions for whole-body PET reports. Twelve language models were trained on a corpus of PET reports using the teacher-forcing algorithm, with the report findings as input and the clinical impressions as reference. An extra input token encodes the reading physician's identity, allowing models to learn physician-specific reporting styles. Our corpus comprised 37,370 retrospective PET reports collected from our institution between 2010 and 2022. To identify the best LLM, 30 evaluation metrics were benchmarked against quality scores from two nuclear medicine (NM) physicians, with the most aligned metrics selecting the model for expert evaluation. In a subset of data, model-generated impressions and original clinical impressions were assessed by three NM physicians according to 6 quality dimensions (3-point scale) and an overall utility score (5-point scale). Each physician reviewed 12 of their own reports and 12 reports from other physicians. Bootstrap resampling was used for statistical analysis. Of all evaluation metrics, domain-adapted BARTScore and PEGASUSScore showed the highest Spearman's rank correlations (0.568 and 0.563) with physician preferences. Based on these metrics, the fine-tuned PEGASUS model was selected as the top LLM. When physicians reviewed PEGASUS-generated impressions in their own style, 89% were considered clinically acceptable, with a mean utility score of 4.08 out of 5. Physicians rated these personalized impressions as comparable in overall utility to the impressions dictated by other physicians (4.03, P=0.41). In conclusion, personalized impressions generated by PEGASUS were clinically useful, highlighting its potential to expedite PET reporting.

MedImageInsight: An Open-Source Embedding Model for General Domain Medical Imaging

Noel C. F. Codella, Ying Jin, Shrey Jain, Yu Gu, Ho Hin Lee, Asma Ben Abacha, Alberto Santamaria-Pang, Will Guyman, Naiteek Sangani, Sheng Zhang, Hoifung Poon, Stephanie Hyland, Shruthi Bannur, Javier Alvarez-Valle, Xue Li, John Garrett, Alan McMillan, Gaurav Rajguru, Madhu Maddi, Nilesh Vijayrania, Rehaan Bhimai, Nick Mecklenburg, Rupal Jain, Daniel Holstein, Naveen Gaur, Vijay Aski, Jenq-Neng Hwang, Thomas Lin, Ivan Tarapov, Matthew Lungren, Mu Wei

In this work, we present MedImageInsight, an open-source medical imaging embedding model. MedImageInsight is trained on medical images with associated text and labels across a diverse collection of domains, including X-Ray, CT, MRI, dermoscopy, OCT, fundus photography, ultrasound, histopathology, and mammography. Rigorous evaluations demonstrate MedImageInsight's ability to achieve state-of-the-art (SOTA) or human expert level performance across classification, image-image search, and fine-tuning tasks. Specifically, on public datasets, MedImageInsight achieves SOTA in CT 3D medical image retrieval, as well as SOTA in disease classification and search for chest X-ray, dermatology, and OCT imaging. Furthermore, MedImageInsight achieves human expert performance in bone age estimation (on both public and partner data), as well as AUC above 0.9 in most other domains. When paired with a text decoder, MedImageInsight achieves near SOTA level single image report findings generation with less than 10\% the parameters of other models. Compared to fine-tuning GPT-4o with only MIMIC-CXR data for the same task, MedImageInsight outperforms in clinical metrics, but underperforms on lexical metrics where GPT-4o sets a new SOTA. Importantly for regulatory purposes, MedImageInsight can generate ROC curves, adjust sensitivity and specificity based on clinical need, and provide evidence-based decision support through image-image search (which can also enable retrieval augmented generation). In an independent clinical evaluation of image-image search in chest X-ray, MedImageInsight outperformed every other publicly available foundation model evaluated by large margins (over 6 points AUC), and significantly outperformed other models in terms of AI fairness (across age and gender). We hope releasing MedImageInsight will help enhance collective progress in medical imaging AI research and development.

Federated Learning Enables Big Data for Rare Cancer Boundary Detection

Sarthak Pati, Ujjwal Baid, Brandon Edwards, Micah Sheller, Shih-Han Wang, G Anthony Reina, Patrick Foley, Alexey Gruzdev, Deepthi Karkada, Christos Davatzikos, Chiharu Sako, Satyam Ghodasara, Michel Bilello, Suyash Mohan, Philipp Vollmuth, Gianluca Brugnara, Chandrakanth J Preetha, Felix Sahm, Klaus Maier-Hein, Maximilian Zenk, Martin Bendszus, Wolfgang Wick, Evan Calabrese, Jeffrey Rudie, Javier Villanueva-Meyer, Soonmee Cha, Madhura Ingalhalikar, Manali Jadhav, Umang Pandey, Jitender Saini, John Garrett, Matthew Larson, Robert Jeraj, Stuart Currie, Russell Frood, Kavi Fatania, Raymond Y Huang, Ken Chang, Carmen Balana, Jaume Capellades, Josep Puig, Johannes Trenkler, Josef Pichler, Georg Necker, Andreas Haunschmidt, Stephan Meckel, Gaurav Shukla, Spencer Liem, Gregory S Alexander, Joseph Lombardo, Joshua D Palmer, Adam E Flanders, Adam P Dicker, Haris I Sair, Craig K Jones, Archana Venkataraman, Meirui Jiang, Tiffany Y So, Cheng Chen, Pheng Ann Heng, Qi Dou, Michal Kozubek, Filip Lux, Jan Michálek, Petr Matula, Miloš Keřkovský, Tereza Kopřivová, Marek Dostál, Václav Vybíhal, Michael A Vogelbaum, J Ross Mitchell, Joaquim Farinhas, Joseph A Maldjian, Chandan Ganesh Bangalore Yogananda, Marco C Pinho, Divya Reddy, James Holcomb, Benjamin C Wagner, Benjamin M Ellingson, Timothy F Cloughesy, Catalina Raymond, Talia Oughourlian, Akifumi Hagiwara, Chencai Wang, Minh-Son To, Sargam Bhardwaj, Chee Chong, Marc Agzarian, Alexandre Xavier Falcão, Samuel B Martins, Bernardo C A Teixeira, Flávia Sprenger, David Menotti, Diego R Lucio, Pamela LaMontagne, Daniel Marcus, Benedikt Wiestler, Florian Kofler, Ivan Ezhov, Marie Metz, Rajan Jain, Matthew Lee, Yvonne W Lui, Richard McKinley, Johannes Slotboom, Piotr Radojewski, Raphael Meier, Roland Wiest, Derrick Murcia, Eric Fu, Rourke Haas, John Thompson, David Ryan Ormond, Chaitra Badve, Andrew E Sloan, Vachan Vadmal, Kristin Waite, Rivka R Colen, Linmin Pei, Murat Ak, Ashok Srinivasan, J Rajiv Bapuraj, Arvind Rao, Nicholas Wang, Ota Yoshiaki, Toshio Moritani, Sevcan Turk, Joonsang Lee, Snehal Prabhudesai, Fanny Morón, Jacob Mandel, Konstantinos Kamnitsas, Ben Glocker, Luke V M Dixon, Matthew Williams, Peter Zampakis, Vasileios Panagiotopoulos, Panagiotis Tsiganos, Sotiris Alexiou, Ilias Haliassos, Evangelia I Zacharaki, Konstantinos Moustakas, Christina Kalogeropoulou, Dimitrios M Kardamakis, Yoon Seong Choi, Seung-Koo Lee, Jong Hee Chang, Sung Soo Ahn, Bing Luo, Laila Poisson, Ning Wen, Pallavi Tiwari, Ruchika Verma, Rohan Bareja, Ipsa Yadav, Jonathan Chen, Neeraj Kumar, Marion Smits, Sebastian R van der Voort, Ahmed Alafandi, Fatih Incekara, Maarten MJ Wijnenga, Georgios Kapsas, Renske Gahrmann, Joost W Schouten, Hendrikus J Dubbink, Arnaud JPE Vincent, Martin J van den Bent, Pim J French, Stefan Klein, Yading Yuan, Sonam Sharma, Tzu-Chi Tseng, Saba Adabi, Simone P Niclou, Olivier Keunen, Ann-Christin Hau, Martin Vallières, David Fortin, Martin Lepage, Bennett Landman, Karthik Ramadass, Kaiwen Xu, Silky Chotai, Lola B Chambless, Akshitkumar Mistry, Reid C Thompson, Yuriy Gusev, Krithika Bhuvaneshwar, Anousheh Sayah, Camelia Bencheqroun, Anas Belouali, Subha Madhavan, Thomas C Booth, Alysha Chelliah, Marc Modat, Haris Shuaib, Carmen Dragos, Aly Abayazeed, Kenneth Kolodziej, Michael Hill, Ahmed Abbassy, Shady Gamal, Mahmoud Mekhaimar, Mohamed Qayati, Mauricio Reyes, Ji Eun Park, Jihye Yun, Ho Sung Kim, Abhishek Mahajan, Mark Muzi, Sean Benson, Regina G H Beets-Tan, Jonas Teuwen, Alejandro Herrera-Trujillo, Maria Trujillo, William Escobar, Ana Abello, Jose Bernal, Jhon Gómez, Joseph Choi, Stephen Baek, Yusung Kim, Heba Ismael, Bryan Allen, John M Buatti, Aikaterini Kotrotsou, Hongwei Li, Tobias Weiss, Michael Weller, Andrea Bink, Bertrand Pouymayou, Hassan F Shaykh, Joel Saltz, Prateek Prasanna, Sampurna Shrestha, Kartik M Mani, David Payne, Tahsin Kurc, Enrique Pelaez, Heydy Franco-Maldonado, Francis Loayza, Sebastian Quevedo, Pamela Guevara, Esteban Torche, Cristobal Mendoza, Franco Vera, Elvis Ríos, Eduardo López, Sergio A Velastin, Godwin Ogbole, Dotun Oyekunle, Olubunmi Odafe-Oyibotha, Babatunde Osobu, Mustapha Shu'aibu, Adeleye Dorcas, Mayowa Soneye, Farouk Dako, Amber L Simpson, Mohammad Hamghalam, Jacob J Peoples, Ricky Hu, Anh Tran, Danielle Cutler, Fabio Y Moraes, Michael A Boss, James Gimpel, Deepak Kattil Veettil, Kendall Schmidt, Brian Bialecki, Sailaja Marella, Cynthia Price, Lisa Cimino, Charles Apgar, Prashant Shah, Bjoern Menze, Jill S Barnholtz-Sloan, Jason Martin, Spyridon Bakas

Knowledge-grounded Adaptation Strategy for Vision-language Models: Building Unique Case-set for Screening Mammograms for Residents Training

Aisha Urooj Khan, John Garrett, Tyler Bradshaw, Lonie Salkowski, Jiwoong Jason Jeong, Amara Tariq, Imon Banerjee

A visual-language model (VLM) pre-trained on natural images and text pairs poses a significant barrier when applied to medical contexts due to domain shift. Yet, adapting or fine-tuning these VLMs for medical use presents considerable hurdles, including domain misalignment, limited access to extensive datasets, and high-class imbalances. Hence, there is a pressing need for strategies to effectively adapt these VLMs to the medical domain, as such adaptations would prove immensely valuable in healthcare applications. In this study, we propose a framework designed to adeptly tailor VLMs to the medical domain, employing selective sampling and hard-negative mining techniques for enhanced performance in retrieval tasks. We validate the efficacy of our proposed approach by implementing it across two distinct VLMs: the in-domain VLM (MedCLIP) and out-of-domain VLMs (ALBEF). We assess the performance of these models both in their original off-the-shelf state and after undergoing our proposed training strategies, using two extensive datasets containing mammograms and their corresponding reports. Our evaluation spans zero-shot, few-shot, and supervised scenarios. Through our approach, we observe a notable enhancement in Recall@K performance for the image-text retrieval task.

From Embeddings to Accuracy: Comparing Foundation Models for Radiographic Classification

Xue Li, Jameson Merkow, Noel C. F. Codella, Alberto Santamaria-Pang, Naiteek Sangani, Alexander Ersoy, Christopher Burt, John W. Garrett, Richard J. Bruce, Joshua D. Warner, Tyler Bradshaw, Ivan Tarapov, Matthew P. Lungren, Alan B. McMillan

Foundation models provide robust embeddings for diverse tasks, including medical imaging. We evaluate embeddings from seven general and medical-specific foundation models (e.g., DenseNet121, BiomedCLIP, MedImageInsight, Rad-DINO, CXR-Foundation) for training lightweight adapters in multi-class radiography classification. Using a dataset of 8,842 radiographs across seven classes, we trained adapters with algorithms like K-Nearest Neighbors, logistic regression, SVM, random forest, and MLP. The combination of MedImageInsight embeddings with an SVM or MLP adapter achieved the highest mean area under the curve (mAUC) of 93.1%. This performance was statistically superior to other models, including MedSigLIP with an MLP (91.0%), Rad-DINO with an SVM (90.7%), and CXR-Foundation with logistic regression (88.6%). In contrast, models like BiomedCLIP (82.8%) and Med-Flamingo (78.5%) showed lower performance. Crucially, these lightweight adapters are computationally efficient, training in minutes and performing inference in seconds on a CPU, making them practical for clinical use. A fairness analysis of the top-performing MedImageInsight adapter revealed minimal performance disparities across patient gender (within 1.8%) and age groups (std. dev < 1.4%), with no significant statistical differences. These findings confirm that embeddings from specialized foundation models, particularly MedImageInsight, can power accurate, efficient, and equitable diagnostic tools using simple, lightweight adapters.

A Generalizable Artificial Intelligence Model for COVID-19 Classification Task Using Chest X-ray Radiographs: Evaluated Over Four Clinical Datasets with 15,097 Patients

Ran Zhang, Xin Tie, John W. Garrett, Dalton Griner, Zhihua Qi, Nicholas B. Bevins, Scott B. Reeder, Guang-Hong Chen

Purpose: To answer the long-standing question of whether a model trained from a single clinical site can be generalized to external sites. Materials and Methods: 17,537 chest x-ray radiographs (CXRs) from 3,264 COVID-19-positive patients and 4,802 COVID-19-negative patients were collected from a single site for AI model development. The generalizability of the trained model was retrospectively evaluated using four different real-world clinical datasets with a total of 26,633 CXRs from 15,097 patients (3,277 COVID-19-positive patients). The area under the receiver operating characteristic curve (AUC) was used to assess diagnostic performance. Results: The AI model trained using a single-source clinical dataset achieved an AUC of 0.82 (95% CI: 0.80, 0.84) when applied to the internal temporal test set. When applied to datasets from two external clinical sites, an AUC of 0.81 (95% CI: 0.80, 0.82) and 0.82 (95% CI: 0.80, 0.84) were achieved. An AUC of 0.79 (95% CI: 0.77, 0.81) was achieved when applied to a multi-institutional COVID-19 dataset collected by the Medical Imaging and Data Resource Center (MIDRC). A power-law dependence, N^(k )(k is empirically found to be -0.21 to -0.25), indicates a relatively weak performance dependence on the training data sizes. Conclusion: COVID-19 classification AI model trained using well-curated data from a single clinical site is generalizable to external clinical sites without a significant drop in performance.

Feasibility Assessment of an Optically Powered Digital Retinal Prosthesis Architecture for Retinal Ganglion Cell Stimulation

William Lemaire, Maher Benhouria, Konin Koua, Wei Tong, Gabriel Martin-Hardy, Melanie Stamp, Kumaravelu Ganesan, Louis-Philippe Gauthier, Marwan Besrour, Arman Ahnood, David John Garrett, Sébastien Roy, Michael Ibbotson, Steven Prawer, Réjean Fontaine

Clinical trials previously demonstrated the notable capacity to elicit visual percepts in blind patients affected with retinal diseases by electrically stimulating the remaining neurons on the retina. However, these implants restored very limited visual acuity and required transcutaneous cables traversing the eyeball, leading to reduced reliability and complex surgery with high postoperative infection risks. To overcome the limitations imposed by cables, a retinal implant architecture in which near-infrared illumination carries both power and data through the pupil to a digital stimulation controller is presented. A high efficiency multi-junction photovoltaic cell transduces the optical power to a CMOS stimulator capable of delivering flexible interleaved sequential stimulation through a diamond microelectrode array. To demonstrate the capacity to elicit a neural response with this approach while complying with the optical irradiance limit at the pupil, fluorescence imaging with a calcium indicator is used on a degenerate rat retina. The power delivered by the laser at the permissible irradiance of 4 mW/mm2 at 850 nm is shown to be sufficient to both power the stimulator ASIC and elicit a response in retinal ganglion cells (RGCs), with the ability to generate of up to 35 000 pulses per second at the average stimulation threshold. This confirms the feasibility of generating a response in RGCs with an infrared-powered digital architecture capable of delivering complex sequential stimulation patterns at high repetition rates, albeit with some limitations.