Research map / Architectures and efficiency / CNN design
ParseCaps: An Interpretable Parsing Capsule Network for Medical Image Diagnosis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ParseCaps, a deep capsule network with parse-tree-like structure (PConvCaps layers, sparse axial attention routing) and concept losses that makes each global capsule dimension map to a human-understandable concept, evaluated on brain MRI and skin lesion classification.
Problem
Capsule networks offer interpretability via hierarchical part-whole structure, but deep capsule networks lose parse-tree-like hierarchy and their instantiation parameters lack clear concept meaning; medical image classification needs interpretable models.
Approach
Combine a parse-convolutional capsule layer that reduces capsule count while increasing dimension with sparse axial attention routing (alpha-entmax over child-parent couplings plus attention over instantiation dimensions), and add presence/triplet or reconstruction losses so global capsule dimensions align with concepts.
Key results
- ParseCaps reaches 99.38% accuracy on CE-MRI, above DeepCaps (93.69%) and the best CNN baseline (94.90%).
- On affNIST, ParseCaps achieves 84.32% accuracy versus 79.00% for the baseline capsule network and 66.00% for a CNN.
- On PH2, ParseCaps reaches 97.53% accuracy, and the concept loss combination lowers explanation error to 0.98 from 4.02 with classification loss alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ParseCaps instead of Modified GoogleNet | CE-MRI classification · CE-MRI | accuracy: 94.90 -> 99.38 | Helped |
| use ParseCaps instead of Coherent CNN | skin-image classification · PH² | accuracy: 96.00 -> 97.53 | Helped |
| use ParseCaps instead of CCNN | skin-image classification · D7 | accuracy: 84.27 -> 83.42 | Hurt |
| use ParseCaps instead of Coherent CNN | skin-image classification · PH²D7 | accuracy: 84.44 -> 87.96 | Helped |
| add triplet loss L_t (ablation) | ParseCaps with L_c + L_p · PH² | ACC: 95.01 -> 97.53 | Helped |
| add triplet loss L_t (ablation) | ParseCaps with L_c + L_p · PH² | EE: 3.93 -> 0.98 | Helped |
| add presentation loss L_p (ablation) | ParseCaps with L_c · PH² | ACC: 94.89 -> 95.01 | Helped |
| add presentation loss L_p (ablation) | ParseCaps with L_c · PH² | EE: 4.02 -> 3.93 | Helped |
| use SAA routing instead of Attention routing (ablation) | 3-layer CapsNet · MNIST | ACC: 98.23 -> 98.66 | Helped |
| use SAA routing instead of Attention routing (ablation) | 3-layer CapsNet · MNIST | FLOPS: 10627M -> 279M | Helped |
| use SAA routing instead of Attention routing (ablation) | 3-layer CapsNet · MNIST | FPS: 277.55 -> 387.88 | Helped |
| add the parse-tree-like structure | CapsNet with SAA routing · affNIST | accuracy: 79.00 -> 84.32 | Helped |
Surprising findings
- Removing the presentation loss alone barely changes PH2 accuracy (94.89 to 95.01) while the triplet loss yields the main gain in explanation error.
- Dynamic routing gave a lower accuracy (98.35%) than SAA routing (99.01%) while being far more costly in FLOPS.
Limitations
- The parse-tree-like structure is not a strict tree since child capsules can connect to multiple parents.
- Concept interpretability without concept labels is shown only on ImageNet-mini, not on medical data, because prototype meanings are hard for non-experts to annotate.
Experimental setup
Datasets: CE-MRI PH2 Derm7pt MNIST affNIST ImageNet-mini
Benchmarks: CE-MRI PH2 Derm7pt MNIST affNIST ImageNet-mini
Metrics: accuracy average precision average recall average F1 specificity explanation error FPS FLOPS
Baselines: DeepCaps OrthCaps CapsNet BrainCaps Modified GoogleNet Block-wise VGG19 Coherent CNN Skin VGGNet CCNN SqueezeNet
Builds on: dynamic routing capsule networks DeepCaps OrthCaps alpha-entmax
Closest papers by meaning
- Why Capsule Neural Networks Do Not Scale: Challenging the Dynamic Parse-Tree Assumption (AAAI 2023 · similarity 0.88)
- STAR-Caps: Capsule Networks with Straight-Through Attentive Routing (NeurIPS 2019 · similarity 0.86)
- HP-Capsule: Unsupervised Face Part Discovery by Hierarchical Parsing Capsule Network (CVPR 2022 · similarity 0.85)
- Interpretable Graph Capsule Networks for Object Recognition (AAAI 2021 · similarity 0.85)
- OrthCaps: An Orthogonal CapsNet with Sparse Attention Routing and Pruning (CVPR 2024 · similarity 0.84)
- Towards Multi-dimensional Explanation Alignment for Medical Classification (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.