MOXIE: Discovering Alternative Explanations for Biomedical Image Classifiers

Abiha Tahsin Chowdhury; Rahul Dubey

cs.CVOct 3, 2026

MOXIE: Discovering Alternative Explanations for Biomedical Image Classifiers

Authors: Abiha Tahsin Chowdhury, Rahul Dubey

Organizations: Department of Computer Science, Missouri State University, Springfield, MO, USA

Abstract

Segment-based explanation methods such as LIME return a single explanation for each prediction, computed from one fixed image segmentation. This hides two important facts: a prediction can be supported by many different sets of image segments, and the segmentation itself shapes which explanations can be found. We introduce MOXIE (Multi-Objective eXplanation Imaging Engine), an evolutionary framework that searches for segment subsets that preserve the classifier's confidence while keeping as little of the image as possible. Instead of one explanation, MOXIE returns a Pareto front of alternative explanations that range from compact to highly faithful. We evaluate MOXIE with NSGA-II and four segmentation methods (SLIC, Felzenszwalb, Watershed and Voronoi) on BloodMNIST and HAM10000 datasets, using the same evaluation budget as LIME. Results show that MOXIE achieves a higher hypervolume than LIME on every image. LIME's explanations often appear convincing, yet the classifier's confidence collapses when only the highlighted segments are shown. MOXIE's fronts reveal how much of the image is needed to preserve the model's confidence and which contextual regions influence it. We also find that segmentation strongly affects evaluation: methods with unequal segment sizes appear most compact when segments are counted. These results show that alternative explanations provide a more complete view of a model's decision than a single explanation.

Figures & tables

Fig. 1: Overview of MOXIE. The input image is divided into segments by one of four segmentation methods. Each candidate explanation is a genome with one gene per segment; segments whose gene exceeds $\tau=0.7$ are kept and the rest are set to black.

Panel of Figure 2 — Fig. 2: A bloodMNIST image segmentation example with Voronoi, SLIC, Watershed and Felzenszwalb methods. Black segments are excluded from the explanation. Each segmentation method produces different segments, and therefore a different trade-off between faithfulness and compactness.

Panel of Figure 3 — Fig. 3: Pareto fronts from the four segmentation methods and the LIME solution for a representative image from each dataset.

Fig. 4: Comparison of explanations obtained using different segmentation methods. Red indicates the five largest retained segments by area. Watershed and Felzenszwalb retain fewer but larger segments, covering a greater portion of the image than SLIC and Voronoi in these examples.

Dataset	Modality	Classes	Test acc.
BloodMNIST	Blood-cell	8	97.87%
HAM10000	Dermoscopy	7	84.88%

TABLE I: Classifier performance where DenseNet121 is fine-tuned for 15 epochs at

224\times 224

image size.

Dataset	Segmentation	$n$	MOXIE HV	LIME HV	Wins	LIME dom.	$p_{\text{Holm}}$
BloodMNIST	SLIC	50	0.512 $\pm$ 0.298	0.107 $\pm$ 0.265	50/50	29%	$1.4\times 10^{-14}$
	Felzenszwalb	50	0.669 $\pm$ 0.262	0.122 $\pm$ 0.269	50/50	36%	$1.4\times 10^{-14}$
	Watershed	50	0.675 $\pm$ 0.259	0.121 $\pm$ 0.271	50/50	39%	$1.4\times 10^{-14}$
	Voronoi	50	0.554 $\pm$ 0.281	0.108 $\pm$ 0.266	50/50	27%	$1.4\times 10^{-14}$
HAM10000	SLIC	50	0.933 $\pm$ 0.139	0.750 $\pm$ 0.269	50/50	95%	$1.4\times 10^{-14}$
	Felzenszwalb	50	0.933 $\pm$ 0.160	0.777 $\pm$ 0.234	50/50	97%	$1.4\times 10^{-14}$

TABLE II: Quantitative comparison of MOXIE and LIME under a matched budget of 2,500 model evaluations per run. HV is reported as mean ± SD across images, with each image averaged over 10 independent runs. “Wins” denotes the number of images for which MOXIE achieves higher HV than LIME, and “LIME dom.” denotes the percentage of LIME solutions dominated by the MOXIE Pareto front.

p_{\mathrm{Holm}}

reports Holm-corrected statistical significance.

Fig. 5: Distribution of hypervolume across 50 test images for MOXIE and LIME using SLIC, Felzenszwalb, Watershed, and Voronoi segmentation on BloodMNIST and HAM10000. Each value represents the mean HV over 10 independent runs for an image. Higher HV indicates a better fidelity–compactness trade-off.

Fig. 6: Pareto-optimal explanations found by MOXIE for a HAM10000 basal cell carcinoma image (confidence 0.99), compared with LIME. Black regions indicate removed segments; percentages denote retained segments and $f_{1}$ denotes confidence deviation.

Fig. 7: Comparison of three Pareto-optimal MOXIE explanations (A–C) from Fig. 6 , with predicted-class confidence shown for each. Green and red indicate regions added and removed, respectively, relative to the preceding explanation; the composite shows where regions first appear.

Explainable Image Classification

Explainable image classification develops ways to show which pixels, regions, concepts, or model features support an image classifier’s predictions.

See all 91 papers

When Superpixels Fail on Documents: A Study of Segmentation for LIME Explanations
Sep 7, 2026Quentin Telnoff, Emanuela Boros, Mickaël Coustaty +3Explainable Artificial Intelligence Explainable Image Classification
1L3i, University of La Rochelle · 2Itesoft
Explanation-Aware Learning for Enhanced Interpretability in Biomedical Imaging
May 11, 2026Zubair Faruqui, Rahul DubeyMedical Image Classification Explainable Medical Image Analysis
Department of Computer Science, Missouri State University
Why Does It Look There? Structured Explanations for Image Classification
Mar 10, 2026Jiarui Li, Zixiang Yin, Samuel J Landry +2Explainable Image Classification Neural Network Interpretability
Department of Computer Science, Tulane University, New Orleans LA 70118, USA · Department of Biochemistry and Molecular Biology, Tulane University School of Medicine, New Orleans LA 70112, USA

Sep 7, 2026J/K move · Enter open · S save

When Superpixels Fail on Documents: A Study of Segmentation for LIME Explanations

Quentin Telnoff, Emanuela Boros, Mickaël Coustaty, Robin Jarry+2

1L3i, University of La Rochelle · 2Itesoft

Explainable Artificial Intelligence Explainable Image Classification Document Image Understanding

Open paper AskPDF arXiv

Post-hoc explanation methods are widely used to inspect image classifiers, but their reliability depends on design choices that are often treated as implementation details. We study this issue for LIME on document image classification, focusing on the segmentation step that defines the interpretable units being perturbed. Standard image-based LIME typically relies on natural-image superpixels, which are poorly aligned with document structure such as text regions, layout blocks, and identification codes. Using RVL-CDIP, we compare Quickshift and SLIC with document-aware segmentations based on OCR bounding boxes and regular grids. Our results show that segmentation strongly affects explanation consistency, correctness, and local fidelity. Document-aware segmentations produce more stable and faithful explanations, require fewer perturbations to converge, and expose shortcut behaviour based on document identification codes, a known RVL-CDIP bias that superpixel-based LIME often obscures. These findings show that reliable post-hoc explanation requires domain-aware interpretable representations, and that segmentation should be treated as part of the explanation method rather than as neutral preprocessing.

Abstract

Explainable Image Classification

Explore similar work

When Superpixels Fail on Documents: A Study of Segmentation for LIME Explanations

Explanation-Aware Learning for Enhanced Interpretability in Biomedical Imaging

Why Does It Look There? Structured Explanations for Image Classification