Technical Summary: MammoExpert

Problem Statement

Breast cancer remains a leading cause of mortality, with mammography serving as a critical diagnostic tool. While AI systems have achieved success in mammography-based diagnosis, they predominantly rely on pattern recognition, lacking the ability to perform the probabilistic diagnostic reasoning characteristic of human experts. This limitation restricts the interpretability of human-AI interactions and hinders the analysis of complex or ambiguous cases. Furthermore, existing publicly available mammography datasets (e.g., CBIS-DDSM, VinDr-Mammo) are often limited in scale, lack rich pathological annotations (specifically WHO-classified subtypes), and do not include structured diagnostic reasoning data. Consequently, AI models cannot learn the intermediate reasoning steps required for trustworthy clinical decision-making.

Methodology

The paper introduces MammoExpert, a novel benchmark dataset and framework designed to bridge the gap between image-based pattern recognition and explicit Chain-of-Thought (CoT) reasoning in mammography.

1. Dataset Construction (MammoExpert):

• Scale and Scope: The dataset comprises 2,379 mammography images covering 1,304 cases.
• Pathological Granularity: It includes 67 WHO-classified histopathological subtypes (30 benign and 37 malignant), addressing the scarcity of fine-grained subtype data in public repositories.
• Annotation Process: Nine senior radiologists (with 15 ± 6 years of experience) annotated each case through a three-phase workflow:
  • Primal Observation: Detection of lesions and calcifications, including location and presence.
  • Factual Assessment: Detailed analysis of lesion characteristics (shape, margin, density) and calcification patterns.
  • Diagnostic Synthesis: Assignment of BI-RADS scores and mapping to specific cancer subtypes.
• Data Structure: Each case includes 42 radiographic features, structured text captions, and ICD codes. The dataset is split into 1,044 training cases and 260 test cases, with careful balancing of subtype distributions.

2. Chain-of-Thought (CoT) Framework:

• The authors propose a three-stage reasoning process derived from clinical workflows: (i) primal observation, (ii) factual assessment, and (iii) diagnostic synthesis.
• Model Training: A Vision-Language Model (VLM), specifically Qwen2.5-VL-3B, is fine-tuned using Supervised Fine-Tuning (SFT). The training data is structured to guide the model through the three-phase analytical procedure explicitly.
• Prompting Strategy: The model is prompted to "analyze the mammography image step by step," generating a <thought> section that details the observation, feature analysis, and reasoning before outputting a final diagnosis (Benign/Malignant) and BI-RADS score.

3. Evaluation:

• Tasks: Binary malignancy classification and multi-subtype classification.
• Baselines: The study compares the proposed CoT approach against standard CNNs (ResNet variants), Vision Transformers (Swin-Transformer), and existing Vision-Language models (MedCLIP, BioViL, PLIP, etc.) trained on public datasets like CBIS-DDSM, INBreast, and VinDr-Mammo.
• Metrics: Accuracy (ACC), AUROC, AUPRC, and Recall, calculated with 95% confidence intervals via bootstrap resampling.

Key Contributions

1. MammoExpert Dataset: The creation of the first open-source mammography dataset featuring comprehensive CoT reasoning annotations across 67 WHO-classified histopathological subtypes.
2. Three-Stage Reasoning Framework: The proposal and validation of a diagnostic workflow (Observation → Assessment → Synthesis) that mimics expert radiologist logic, demonstrating that explicit reasoning improves model performance over direct end-to-end prediction.
3. Proof-of-Concept for Reasoning: Empirical evidence showing that AI systems trained with CoT annotations achieve superior accuracy and interpretability, particularly in distinguishing morphologically similar but pathologically distinct subtypes and handling rare cases.

Experimental Results

The paper reports significant performance gains across multiple datasets when integrating MammoExpert and the CoT reasoning framework:

• CBIS-DDSM: Combining MammoExpert with the Qwen2.5-VL model improved classification accuracy from 70.90% (baseline on CBIS-DDSM alone) to 78.04%, a 7.1% absolute gain. The CoT reasoning component alone contributed an additional 4% gain on the MammoExpert test set compared to direct prediction.
• INBreast: Accuracy improved from 75.52% to 82.44% with the full approach (MammoExpert + CoT), representing a 6.9% gain.
• VinDr-Mammo: Accuracy increased from 78.62% to 85.32%, a 6.7% improvement.
• Rare Subtypes: The model demonstrated stable performance on rare subtypes (n ≤ 20), achieving accuracies between 61% and 66%. CoT reasoning consistently improved Recall and Accuracy for difficult, morphologically similar pairs (e.g., mucinous carcinoma vs. invasive ductal carcinoma).
• Ablation: The CoT-enabled model (Qwen2.5-VL-3B + CoT) achieved 75.68% accuracy on the MammoExpert test set, outperforming the image-only baseline (71.62%) and standard CNNs.

Significance and Claims

The authors claim that MammoExpert serves as a critical benchmark for developing interpretable and trustworthy AI systems for breast lesion diagnosis. By explicitly modeling the diagnostic reasoning process, the framework addresses the "black-box" nature of traditional medical AI, fostering greater clinical trust.

The paper emphasizes that the structured annotations enable models to:

• Decompose complex classification problems into manageable sub-decisions.
• Reduce reliance on global visual shortcuts and mitigate over-prediction of high-frequency classes.
• Provide transparent rationales that align with real-world clinical decision logic.

The work positions itself as a step toward "intelligence and reasonability" in healthcare, suggesting that future research should focus on extending this reasoning framework to other medical imaging modalities while maintaining consistency with expert workflows. The authors acknowledge limitations regarding the dataset's scale for extremely rare subtypes but argue that the structured reasoning annotations partially mitigate data scarcity issues by providing complementary supervisory signals.