Technical Summary: PolyDetect

Problem Statement

Colorectal cancer (CRC) is a leading cause of cancer-related mortality, with colonoscopy serving as the primary screening method. While deep learning models have achieved high segmentation accuracy on controlled benchmark datasets (e.g., Kvasir-SEG), their performance degrades significantly in real-world clinical settings. Real-world colonoscopy images frequently suffer from five specific types of degradation: motion blur, additive Gaussian noise, low contrast (particularly in flat/sessile polyps), JPEG compression artifacts, and poor illumination.

Existing state-of-the-art segmentation models (e.g., PraNet, SANet, Polyp-PVT) generally lack mechanisms to explicitly handle these degradations. They typically treat all training samples uniformly, regardless of image quality, and do not incorporate Image Quality Assessment (IQA) into the training or inference pipeline. This leads to a performance gap where models trained on clean benchmarks fail to generalize to degraded, zero-shot datasets like CVC-ColonDB and ETIS-LaribPolypDB.

Methodology: The PolyDetect Framework

The authors propose PolyDetect, a quality-aware hybrid deep learning framework designed to robustly segment polyps in degraded images. The framework integrates five tightly coupled components:

1. Degradation Simulation Module: During training, the model is exposed to a five-type degradation simulator (Gaussian blur, Gaussian noise, low contrast, JPEG compression, and poor illumination) with a 70% application probability. This ensures the training distribution covers a realistic range of image qualities while retaining clean samples.
2. Image Quality Assessment (IQA) Module: A dual-branch module computes a scalar quality score $q \in [0, 1]$ for each image.
   • BRISQUE Branch: Uses classical Blind/Reference-less Image Spatial Quality Evaluator statistics.
   • Lightweight CNN Branch: A small network (~16K parameters) that learns to regress quality features.
   • Fusion: The final score is a 50/50 weighted average of the normalized BRISQUE score and the CNN output.
3. Quality-Score-Gated Adaptive Enhancement Pipeline: Based on the IQA score $q$, the system selectively applies a three-tier enhancement pipeline to avoid over-processing clean images:
   • $q \geq 0.7$ (Good): No enhancement.
   • $0.4 \leq q < 0.7$ (Moderate): Applies CLAHE (contrast) and DnCNN (denoising).
   • $q < 0.4$ (Poor): Applies full pipeline: Retinex (illumination) $\to$ CLAHE $\to$ DnCNN.
   • Note: In the reported experiments, the DnCNN and CNN quality branch were used with random initialization (fixed filters) rather than being jointly trained, serving as a proof-of-concept for the gating mechanism.
4. Hybrid Segmentation Backbone:
   • Encoder: Pyramid Vision Transformer (PVT-v2-B2) pre-trained on ImageNet-1K to capture global context.
   • Decoder: A CNN U-Net-style decoder with skip connections to restore local spatial details.
   • Boundary Refinement Head: A dedicated module trained on morphologically derived 1-pixel edge maps to sharpen polyp boundaries, critical for the HD95 metric.
5. Quality-Weighted Composite Loss: The loss function weights the contribution of each sample by its quality score $q$:
   $$L = \text{mean}[(L_{Dice} + L_{BCE} + \lambda_b \times L_{Boundary}) \times q]$$
   This down-weights degraded samples during training, preventing the model from memorizing noise patterns.

Key Contributions

The paper identifies three specific gaps in current literature and addresses them through the following contributions:

1. Live IQA Integration: The first polyp segmentation framework to embed a differentiable, live IQA signal as a training-time loss modulator and inference-time gating mechanism.
2. Adaptive Enhancement: A novel quality-score-gated pipeline that prevents artifact introduction in high-quality frames by selectively applying enhancement only when necessary.
3. Quality-Weighted Training: A loss weighting scheme that proportionally reduces the gradient contribution of degraded samples.
4. Hybrid Architecture: A combination of PVT-v2 global context and CNN local detail, augmented by a morphological boundary refinement head.
5. Systematic Robustness Benchmark: A per-degradation evaluation protocol testing five specific degradation types, revealing performance drops in existing models that were previously unquantified.

Experimental Results

The model was trained on 900 images from the Kvasir-SEG dataset and evaluated in a zero-shot setting on CVC-ColonDB (380 images) and ETIS-LaribPolypDB (196 images).

• Performance: PolyDetect achieved the highest average Dice score (0.8066) and IoU (0.7262) across all three datasets.
  • Kvasir-SEG: Dice 0.9106 (outperforming U-Net++ by ~11 pp).
  • CVC-ColonDB: Dice 0.7543 (outperforming SANet by ~0.9 pp).
  • ETIS-LaribPolypDB: Dice 0.7548 (outperforming SANet by ~5.8 pp).
• Statistical Significance: The improvements over CNN-based baselines (U-Net, U-Net++) were statistically significant ($p < 10^{-10}$) across all datasets. Improvements over transformer-based baselines (PraNet, SANet) were significant in some zero-shot comparisons (e.g., ETIS vs. SANet, $p=0.037$) but did not survive Holm correction for multiplicity in all cases.
• Robustness: PolyDetect demonstrated the best absolute Dice scores across all five degradation types on CVC-ColonDB.
• Efficiency: The model runs at 53.62 FPS on a Tesla T4 GPU (27.35M parameters, 12.47 GFLOPs), meeting real-time clinical requirements (typically >25 FPS).
• Ablation Studies:
  • The Boundary Refinement Head was the most critical component; its removal caused a 6.47 percentage point drop in Dice.
  • The Adaptive Enhancement pipeline contributed a 2.39 pp drop when removed.
  • The Quality-Weighted Loss contributed a 0.75 pp drop.
  • Removing IQA and enhancement together resulted in a 3.62 pp drop.

Significance and Claims

The paper claims that PolyDetect offers a feasible path toward robust polyp segmentation in real-world, degraded colonoscopy images by unifying quality assessment, gated enhancement, and quality-weighted training into a single pipeline.

• Clinical Relevance: The framework achieves real-time inference speeds and significantly outperforms standard CNN baselines in zero-shot generalization, addressing the "controlled benchmark vs. real-world deployment" gap.
• Modest Claims: The authors explicitly note that while the framework is promising, the current results rely on a single random seed and non-standard evaluation protocols compared to some SOTA methods (which use combined training sets). They state that the advantage over transformer-based baselines (like SANet) is not fully established without multi-seed validation and that the end-to-end latency (including IQA and enhancement steps) requires further validation before clinical deployment.
• Future Work: The paper suggests that future iterations should employ standard protocols, multiple data splits, full-length ablations, and prospective testing with private clinical data to verify these findings.