BAHOP: Similarity-based Basin Hopping for A fast hyper-parameter search in WSI classification

TL;DR

This work addresses the sensitivity of WSI pre-processing to domain shifts in MIL-based histopathology classification. It introduces BAHOP, a PSNR-based Basin Hopping optimization that rapidly searches preprocessing hyperparameters to boost out-of-domain inference while reducing computational cost. The method demonstrates 5–30% accuracy gains and up to roughly 5× speedups across CAMELYON and TCGA datasets and multiple MIL models, by avoiding costly feature extraction for unlikely parameter sets through a similarity-guided pruning strategy. The approach enables practical, center-specific preprocessing tuning without retraining, enhancing robustness and scalability of WSI classification in diverse clinical settings.

Abstract

Pre-processing whole slide images (WSIs) can impact classification performance. Our study shows that using fixed hyper-parameters for pre-processing out-of-domain WSIs can significantly degrade performance. Therefore, it is critical to search domain-specific hyper-parameters during inference. However, searching for an optimal parameter set is time-consuming. To overcome this, we propose BAHOP, a novel Similarity-based Basin Hopping optimization for fast parameter tuning to enhance inference performance on out-of-domain data. The proposed BAHOP achieves 5\% to 30\% improvement in accuracy with $\times5$ times faster on average.

Figures (6)

• Figure 1: The performance of out-of-domain inference varies with preprocessing parameters across various MIL models and datasets. Consequently, we suggest that each specific center within the dataset should adopt its own preprocessing parameters to maximize performance. The original method involves all centers using fixed default preprocessing hyperparameters, whereas the optimal method allows each center to use its own specific preprocessing hyperparameters determined by our proposed BAHOP.
• Figure 2: PSNR-based Basin Hopping hyper-parameter optimization for out-of-domain inference in WSIs. The hyper-parameter optimization for the pre-processing task is a non-convex optimization that contains many large local optima. Our BAHOP is developed for fast search, and the jumping range of each iteration is controlled by the PSNR threshold.
• Figure 3: The inference performance of out-of-domain data varies with preprocessing parameters across various MIL models and datasets.
• Figure 4: All the heatmaps circled by red boxes correspond to the default hyper-parameter and predict wrong, while heatmaps circled by blue boxes correspond to optimal hyperparameters with correct predictions. Fig.A: Hyper-parameter optimization starts from the default parameter, which is the same as the pre-trained model. The default pre-processed WSI drops many patches that get a high attention score in optimal pre-processed WSI. Fig.B, the dataset is TCGA-NSCLC where dropping tissue can improve accuracy. Fig.C: Some hyper-parameters drop the region of interest(ROI) during pre-processing.
• Figure 5: Relationship between PSNR and Accuracy. The red star stands for the reference object. All the hyper-parameters are compared with the reference object to calculate the PSNR.