Interpretable Tile-Based Classification of Paclitaxel Exposure

TL;DR

This work tackles the challenge of classifying paclitaxel exposure from phase-contrast microscopy where full-image models struggle with subtle dose cues. It proposes a tiling-and-aggregation pipeline built on a ResNet-50 backbone and a KNN head, enabling tile-level representations that are aggregated into image-level predictions. The approach achieves state-of-the-art performance on the Taxol dataset, reaching about 0.97 accuracy and surpassing the prior baseline by roughly 20 percentage points, with five-fold cross-validation confirming robustness. Interpretability analyses using Grad-CAM and Score-CAM reveal that the model relies on localized contextual textures rather than cellular morphology, highlighting both a practical performance gain and concerns about robustness under distribution shift. Open-source code is provided to facilitate reproducibility and further investigation into drug-response phenotyping with tile-based methods.

Abstract

Medical image analysis is central to drug discovery and preclinical evaluation, where scalable, objective readouts can accelerate decision-making. We address classification of paclitaxel (Taxol) exposure from phase-contrast microscopy of C6 glioma cells -- a task with subtle dose differences that challenges full-image models. We propose a simple tiling-and-aggregation pipeline that operates on local patches and combines tile outputs into an image label, achieving state-of-the-art accuracy on the benchmark dataset and improving over the published baseline by around 20 percentage points, with trends confirmed by cross-validation. To understand why tiling is effective, we further apply Grad-CAM and Score-CAM and attention analyses, which enhance model interpretability and point toward robustness-oriented directions for future medical image research. Code is released to facilitate reproduction and extension.

Figures (6)

• Figure 1: Representative phase-contrast microscopy images of C6 glioma cells from each experimental group.
• Figure 2: Overall pipeline for paclitaxel exposure classification. Phase 1 tiles the image into an r-by-c grid. Phase 2 classifies each tile with a shared CNN and a K–NN head. Phase 3 aggregates tile outputs by majority or probability voting to obtain the image-level label.
• Figure 3: Example of tiling: 100 nM class, image 068, divided into 6 rows by 7 columns.
• Figure 4: Accuracy across tiling configurations.
• Figure 5: Example tiles and Grad-CAM maps from the 100 nM class (Image 52). (a) Tile 30 (no visible cells). (b) Grad-CAM for (a). (c) Tile 09 (visible cells). (d) Grad-CAM for (c). In both cases, the most salient regions emphasize contextual textures rather than cellular morphology.