Abstract
Assessing the difficulty of Total Mesorectal Excision (TME) for rectal cancer is critical for surgical planning but currently relies on subjective clinical judgment. This paper introduces DRIMV_TSK, a novel framework that integrates incomplete multi-view data (MRI and clinical records) to objectively evaluate surgical difficulty. By combining dual-representation learning with a Takagi-Sugeno-Kang (TSK) fuzzy system, the model addresses data incompleteness and the "black box" nature of deep learning. Experimental results on the MVCR dataset show that DRIMV_TSK achieves state-of-the-art accuracy (0.9471) and maintains robustness even with 70% missing data, while providing interpretable rule-based logic for clinical decision support.
Introduction
Total mesorectal excision (TME) is the standard treatment for rectal cancer, yet its complexity varies significantly based on individual patient anatomy and tumor characteristics. Objective preoperative assessment is vital but challenging due to two primary issues in clinical AI: data incompleteness and lack of interpretability. Patients often lack a complete set of high-resolution MRI scans and clinical metrics, and standard "black box" models fail to provide the transparency required for medical trust.
To address these challenges, we present the Multi-View Rectal Cancer (MVCR) dataset, comprising 151 patients with high-resolution MRI, pressed-fat MRI, and clinical data. We propose DRIMV_TSK to leverage this multi-view information effectively, even when views are missing.
Methods
The DRIMV_TSK framework operates in two primary stages:
Stage 1: Dual Representation Incomplete Multi-view Learning (IMV_DRL)
Unlike traditional models that only seek shared information, IMV_DRL extracts both common and view-specific representations. It incorporates missing view imputation directly into the learning process, using second-order similarity constraints to ensure that patient relationships in the latent space remain consistent across different data views.
Stage 2: Interpretable TSK Fuzzy System
The extracted features are processed by a TSK fuzzy system, which generates human-readable "If-Then" rules. A cooperative learning mechanism aligns outputs across views, while Shannon entropy automatically weights the reliability of each view (e.g., prioritizing MRI over clinical data when appropriate) for a specific prediction.
The implementation details of the DRIMV_TSK algorithm are summarized in Algorithm 2, which coordinates the learning of imputed data and fuzzy parameters.
Results and Discussion
DRIMV_TSK was evaluated against state-of-the-art algorithms across varying missing data rates (10% to 70%).
- Performance and Robustness: At 10% missing data, the model achieved an Accuracy of 0.9471 and an AUC of 0.9934. Remarkably, even with 70% of data views missing, it maintained an Accuracy of 0.9306, significantly outperforming traditional imputation methods like KNN or mean imputation.
- Ablation Insights: Removing view-specific representations led to a significant performance drop, confirming that unique details in specific MRI sequences are as critical as shared features.
- Interpretability: The model successfully generated linguistic rules (e.g., "The Pelvis Depth is Medium") that align with clinical observations, bridging the gap between AI predictions and surgical intuition.
Conclusion
DRIMV_TSK provides a robust and transparent solution for surgical difficulty assessment in rectal cancer. While currently limited by linear representation learning and a focus on missing views rather than individual features, it sets a new benchmark for handling incomplete medical data. Future work will explore non-linear deep learning integrations and semi-supervised learning to utilize larger, unlabeled clinical datasets.
Summary updated