Yapay Zeka · 8 dk okuma · 22 Nisan 2026
Concept Bottleneck Models Hit Hard Ceiling in Dermoscopy Data
Rough-set analysis reveals 16% of concept profiles in Derm7pt are internally inconsistent, capping model accuracy at 92% regardless of architecture.
Concept Bottleneck Models in dermoscopy face a hard accuracy ceiling due to inherent dataset inconsistencies that no training method can overcome.
- — 50 of 305 unique concept profiles in Derm7pt contain conflicting diagnosis labels, affecting 30% of images.
- — This inconsistency creates a theoretical accuracy ceiling of 92.1% for any hard-concept CBM.
- — Rough set theory identifies boundary-region images as primary sources of concept-level conflict.
- — Symmetric filtering removes all ambiguous cases, producing Derm7pt+ with 705 fully consistent images.
- — EfficientNet-B5 achieves 0.90 label accuracy and 0.70 concept accuracy on the filtered benchmark.
- — Filtering strategy choice (symmetric vs. asymmetric) materially affects both model performance and interpretability.
- — Results establish reproducible baselines for concept-consistent evaluation in medical image classification.
Sık sorulanlar
- A Concept Bottleneck Model (CBM) is a neural network that makes predictions by first predicting human-interpretable concepts (e.g., skin lesion features), then using those concepts to predict the final diagnosis. This design improves interpretability because clinicians can see which concepts the model relied on. However, CBMs are only as accurate as the consistency of their concept annotations in the training data.