The Need to Move Beyond Explainability Toward Chain-of-Thought Reasoning: A Focus on AI for Mammography

Yalda Zafari, Shahd Soliman, Essam A. Rahed, Mohamed Mabrok
Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare, PMLR 317:322-329, 2026.

Abstract

Mammography is one of the primary imaging modalities for breast cancer screening and diagnosis, playing a pivotal role in early detection and mortality reduction. To alleviate the burden on radiologists interpreting mammographic data, artificial intelligence-based models have emerged as decision-making assistants, demonstrating promising results in several studies. However, a critical gap remains between AI capabilities and clinical integration. Current AI systems predominantly employ end-to-end classification approaches that bypass the structured, multi-step reasoning radiologists use in practice. Radiologists systematically detect abnormalities, characterize their features using standardized descriptors, correlate findings across imaging views, perform temporal comparisons, assess information sufficiency, and synthesize evidence into risk-stratified recommendations. In contrast, most AI models map directly from images to diagnoses without transparent intermediate reasoning, limiting their interpretability, clinical utility, and ability to generalize beyond training distributions. This paper examines the radiological chain-of-thought process in mammography interpretation, reviews current AI approaches and their limitations, and proposes a multi-stage reasoning framework that explicitly models each step of clinical decision-making. By decomposing the diagnostic task into sequential reasoning-based stages, this framework aims to create AI systems that not only predict outcomes but also reason transparently in alignment with clinical workflow. Crucially, we explain how contextual information such as breast density serves as a conditioning variable that dynamically optimizes subsequent reasoning stages, mirroring the adaptive decision-making process radiologists employ in clinical practice. We discuss the implications of this approach and identify critical dataset limitations that must be addressed to enable the development of truly reasoning-aware AI in breast imaging.

Cite this Paper

BibTeX
@InProceedings{pmlr-v317-zafari26a, title = {The Need to Move Beyond Explainability Toward Chain-of-Thought Reasoning: A Focus on AI for Mammography}, author = {Zafari, Yalda and Soliman, Shahd and Rahed, Essam A. and Mabrok, Mohamed}, booktitle = {Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare}, pages = {322--329}, year = {2026}, editor = {Wu, Junde and Pan, Jiazhen and Zhu, Jiayuan and Luo, Luyang and Li, Yitong and Xu, Min and Jin, Yueming and Rueckert, Daniel}, volume = {317}, series = {Proceedings of Machine Learning Research}, month = {20--21 Jan}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v317/main/assets/zafari26a/zafari26a.pdf}, url = {https://proceedings.mlr.press/v317/zafari26a.html}, abstract = {Mammography is one of the primary imaging modalities for breast cancer screening and diagnosis, playing a pivotal role in early detection and mortality reduction. To alleviate the burden on radiologists interpreting mammographic data, artificial intelligence-based models have emerged as decision-making assistants, demonstrating promising results in several studies. However, a critical gap remains between AI capabilities and clinical integration. Current AI systems predominantly employ end-to-end classification approaches that bypass the structured, multi-step reasoning radiologists use in practice. Radiologists systematically detect abnormalities, characterize their features using standardized descriptors, correlate findings across imaging views, perform temporal comparisons, assess information sufficiency, and synthesize evidence into risk-stratified recommendations. In contrast, most AI models map directly from images to diagnoses without transparent intermediate reasoning, limiting their interpretability, clinical utility, and ability to generalize beyond training distributions. This paper examines the radiological chain-of-thought process in mammography interpretation, reviews current AI approaches and their limitations, and proposes a multi-stage reasoning framework that explicitly models each step of clinical decision-making. By decomposing the diagnostic task into sequential reasoning-based stages, this framework aims to create AI systems that not only predict outcomes but also reason transparently in alignment with clinical workflow. Crucially, we explain how contextual information such as breast density serves as a conditioning variable that dynamically optimizes subsequent reasoning stages, mirroring the adaptive decision-making process radiologists employ in clinical practice. We discuss the implications of this approach and identify critical dataset limitations that must be addressed to enable the development of truly reasoning-aware AI in breast imaging.} }
Endnote
%0 Conference Paper %T The Need to Move Beyond Explainability Toward Chain-of-Thought Reasoning: A Focus on AI for Mammography %A Yalda Zafari %A Shahd Soliman %A Essam A. Rahed %A Mohamed Mabrok %B Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare %C Proceedings of Machine Learning Research %D 2026 %E Junde Wu %E Jiazhen Pan %E Jiayuan Zhu %E Luyang Luo %E Yitong Li %E Min Xu %E Yueming Jin %E Daniel Rueckert %F pmlr-v317-zafari26a %I PMLR %P 322--329 %U https://proceedings.mlr.press/v317/zafari26a.html %V 317 %X Mammography is one of the primary imaging modalities for breast cancer screening and diagnosis, playing a pivotal role in early detection and mortality reduction. To alleviate the burden on radiologists interpreting mammographic data, artificial intelligence-based models have emerged as decision-making assistants, demonstrating promising results in several studies. However, a critical gap remains between AI capabilities and clinical integration. Current AI systems predominantly employ end-to-end classification approaches that bypass the structured, multi-step reasoning radiologists use in practice. Radiologists systematically detect abnormalities, characterize their features using standardized descriptors, correlate findings across imaging views, perform temporal comparisons, assess information sufficiency, and synthesize evidence into risk-stratified recommendations. In contrast, most AI models map directly from images to diagnoses without transparent intermediate reasoning, limiting their interpretability, clinical utility, and ability to generalize beyond training distributions. This paper examines the radiological chain-of-thought process in mammography interpretation, reviews current AI approaches and their limitations, and proposes a multi-stage reasoning framework that explicitly models each step of clinical decision-making. By decomposing the diagnostic task into sequential reasoning-based stages, this framework aims to create AI systems that not only predict outcomes but also reason transparently in alignment with clinical workflow. Crucially, we explain how contextual information such as breast density serves as a conditioning variable that dynamically optimizes subsequent reasoning stages, mirroring the adaptive decision-making process radiologists employ in clinical practice. We discuss the implications of this approach and identify critical dataset limitations that must be addressed to enable the development of truly reasoning-aware AI in breast imaging.
APA
Zafari, Y., Soliman, S., Rahed, E.A. & Mabrok, M.. (2026). The Need to Move Beyond Explainability Toward Chain-of-Thought Reasoning: A Focus on AI for Mammography. Proceedings of The Second AAAI Bridge Program on AI for Medicine and Healthcare, in Proceedings of Machine Learning Research 317:322-329 Available from https://proceedings.mlr.press/v317/zafari26a.html.

Related Material