288 - Revisiting Robustness: Evaluating CLAHE and Test-Time Augmentation for Catheter Segmentation in X-Ray Fluoroscopy

Purpose: Automated catheter segmentation in fluoroscopy can aid image-guided interventional radiology. The lightweight U-Net with transfer learning proposed by Gherardini et al. (CMPB 2020) demonstrated real-time performance using phantom and in-vivo datasets. Post-processing and augmentation strategies such as contrast-limited adaptive histogram equalization (CLAHE) and test-time augmentation (TTA) are widely used in natural image analysis to enhance robustness. We aimed to evaluate whether CLAHE, TTA, or their combination could improve segmentation performance for catheters in fluoroscopy.

Materials and Methods: We trained a lightweight U-Net on 2000 phantom images with provided labels, then fine-tuned and tested on six in-vivo sequences (T1–T6, total 1207 frames) with binary catheter masks. Four inference strategies were evaluated: baseline (raw images), CLAHE only, TTA only (horizontal/vertical flips), and CLAHE+TTA. Segmentation accuracy was quantified using Dice coefficients, following the same dataset splits as the reference study.

Results: Baseline Dice ranged from 0.13-0.53 across sequences. CLAHE consistently reduced performance (e.g., T1: 0.531 → 0.425; T3: 0.436 → 0.268), indicating intensity distribution mismatch. TTA produced neutral to slightly positive effects, improving Dice in T2 (0.310 → 0.345), T3 (0.436 → 0.471), and T6 (0.168 → 0.184). Combining CLAHE+TTA failed to recover performance, with average Dice decreases of 0.07–0.13 across sequences compared to baseline.

Conclusion: Contrary to expectations, CLAHE preprocessing degraded catheter segmentation accuracy, while simple TTA produced only modest gains. These results highlight that “off-the-shelf” robustness techniques from natural image domains may backfire in fluoroscopy if training and inference distributions are not aligned. For interventional radiology, this emphasizes the importance of modality-specific validation of preprocessing and augmentation strategies before clinical deployment.