Editors: Liron Pantanowitz, MD, PhD, MHA, chair of the Department of Pathology and professor of pathology, University of Pittsburgh Medical Center, and Matthew G. Hanna, MD, vice chair of pathology informatics and associate professor, Department of Pathology, University of Pittsburgh Medical Center.
Predicting MMR deficiency status in endometrial cancer via multiresolution ensemble learning in digital pathology
January 2025—Mismatch repair deficiency is a critical biomarker for identifying patients who may benefit from immunotherapy, and accurate classification is essential to making personalized treatment decisions. The authors conducted a study in which they presented a deep learning-based approach for classifying mismatch repair deficiency (MMR-D) in endometrial cancer using whole slide images of H&E-stained slides. They employed a multiresolution ensemble learning model in which they processed whole slide images at three magnifications—2.5×, 5×, and 10×—using a combination of the deep-learning architectures InceptionResNetV2, EfficientNetB2, and EfficientNetB3. These networks were trained on a data set of 1,168 whole slide images from 325 patients, with each whole slide image labeled by a pathologist for MMR-D or MMR proficiency (MMR-P) based on IHC results for the key MMR proteins MLH1, MSH2, MSH6, and PMS2. The authors addressed color variability in the H&E slides using a CycleGAN-based network for color normalization, ensuring consistency across the data set. They employed a tile-based approach for classification in which individual tiles were classified and aggregated to determine the MMR status of each whole slide image. This approach leveraged the strengths of both low- and high-magnification images. The authors found that the EfficientNetB2 model outperformed the other architectures, achieving an area under the receiver operating characteristic curve of 0.821, with high sensitivity (0.827). This study demonstrated the effectiveness of multiresolution ensemble learning in predicting MMR-D status, with the potential to aid clinical decision-making, particularly in selecting patients for immunotherapy. This method holds promise for improving the accuracy and speed of molecular classification in endometrial cancer, although further validation across diverse data sets is needed to ensure its generalizability. The study results suggest that such artificial intelligence-driven tools could be integrated into clinical workflows to enhance personalized treatment strategies based on molecular characteristics.
Whangbo J, Lee YS, Kim YJ, et al. Predicting mismatch repair deficiency status in endometrial cancer through multi-resolution ensemble learning in digital pathology. J Imaging Inform Med. 2024;37:1674–1682.
Correspondence: Dr. Jisup Kim at jspath@gilhospital.com or Dr. Kwang Gi Kim at kimkg@gachon.ac.kr
Potential risks of using ChatGPT in diagnostic surgical pathology
The application of artificial intelligence has expanded from a traditional, or nongenerative, approach, such as the use of AI algorithms, to also encompass a generative approach, which includes the use of AI tools such as ChatGPT. Generative AI tools can produce original content, such as text, images, code, audio, and video. OpenAI’s popular ChatGPT AI chatbot can generate human-like text in response to user prompts. It likely can be very helpful in pathology, especially when integrated with medical text and tabular databases, as well as with whole slide imaging systems. ChatGPT can be employed for multiple purposes. It can be used in education—for example, to generate case-based questions and provide translation. It can also be used in research for such tasks as literature review and data mining, as well as in administration for document automation and other functions. It may even be used in the clinical service to retrieve guidelines, brainstorm differential diagnoses, and draft explanations of complex pathology reports for laypeople. However, ChatGPT has several limitations. These include privacy concerns when dealing with patient data that contain protected health information; accuracy when performing complex pathology tasks, such as hallucinations; and ethics surrounding such issues as bias and deepfakes. Therefore, more research is required before ChatGPT or equivalent large language models (LLMs) can be used for such diagnostic purposes as drafting pathology reports or analyzing images. The authors herein conducted a study to evaluate the reliability of ChatGPT for addressing pathology-related diagnostic questions in 10 subspecialties and its ability to provide scientific references. The 10 subspecialty areas studied included breast, dermatopathology, gastrointestinal, gynecological, hematopathology, genitourinary, endocrine, neuropathology, bone and soft tissue, and thoracic. For each subspecialty, five clinico-pathological scenarios were created and then presented to ChatGPT as open-ended or multiple-choice questions—for example, What is the most likely differential diagnosis? Scientific references were sometimes requested. The focus of the five scenarios for each subspecialty was randomly chosen by a pathologist using the World Health Organization’s WHO Classification of Tumours series of books as a reference. AI output for the 200 questions generated (five scenarios × 10 subspecialties × two question-type prompts × two reference-type prompts) was evaluated in random fashion by six pathologists. The authors found that ChatGPT delivered helpful answers in 62.2 percent of cases, contained no errors in 32.1 percent of outputs, and had at least one error in the remaining cases. Of the bibliographic references provided by ChatGPT, 70.1 percent were accurate, 12.1 percent were inaccurate, and 17.8 percent were fabricated. The authors concluded that while ChatGPT provided useful responses in almost one-third of their simulated cases, it is not yet adequate for routine diagnostic use in histopathology due to frequency of errors, variability, and imprecise referencing. Therefore, ChatGPT should, for now, serve as a supplementary tool in such areas as pathology education and research and only in preliminary diagnostic tasks.
Guastafierro V, Corbitt DN, Bressan A, et al. Unveiling the risks of ChatGPT in diagnostic surgical pathology. Virchow Archiv. 2024. doi.org/10.1007/s00428-024-03918-1
Correspondence: Dr. Salvatore Lorenzo Renne at salvatore.renne@hunimed.eu