Conference on Medical Imaging with Deep Learning (MIDL) 2026

A Tool Bottleneck Framework for Clinically-Informed
and Interpretable Medical Image Understanding

Christina Liu*1,2Alan Wang*2Joy Hsu2Jiajun Wu2Ehsan Adeli2

1California Institute of Technology    2Stanford University    *Equal Contribution


TBF Overview
Overview of the Tool Bottleneck Framework.

Abstract

Recent tool-use frameworks powered by vision-language models (VLMs) improve image understanding by grounding model predictions with specialized tools. Broadly, these frameworks leverage VLMs and a pre-specified toolbox to decompose the prediction task into multiple tool calls (often deep learning models) which are composed to make a prediction. The dominant approach to composing tools is using text, via function calls embedded in VLM-generated code or natural language. However, these methods often perform poorly on medical image understanding, where salient information is encoded as spatially-localized features that are difficult to compose or fuse via text alone.

To address this, we propose a tool-use framework for medical image understanding called the Tool Bottleneck Framework (TBF), which composes VLM-selected tools using a learned Tool Bottleneck Model (TBM). For a given image and task, TBF leverages an off-the-shelf medical VLM to select tools from a toolbox that each extract clinically-relevant features. Instead of text-based composition, these tools are composed by the TBM, which computes and fuses the tool outputs using a neural network before outputting the final prediction. We propose a simple and effective strategy for TBMs to make predictions with any arbitrary VLM tool selection. Overall, our framework not only improves tool-use in medical imaging contexts, but also yields more interpretable, clinically-grounded predictors. We evaluate TBF on tasks in histopathology and dermatology and find that these advantages enable our framework to perform on par with or better than deep learning-based classifiers, VLMs, and state-of-the-art tool-use frameworks, with particular gains in data-limited regimes.


Method

1. Toolbox. A set of pretrained tools that each extract a specific clinically-relevant feature as a spatial map. For histopathology, tools capture nucleus-level outputs (bounding boxes, centroids, contours, cell types) via HoVer-Net. For dermatology, tools include a lesion segmenter, dermoscopic structure detectors, and color-based malignancy markers.

2. VLM tool selector. MedGemma selects the top-k most relevant tools for the given image and task, prompted with tool descriptions and a task definition — distilling clinical knowledge without additional supervision.

3. Tool Bottleneck Model (TBM). Selected tool outputs are rasterized into multi-channel spatial maps, concatenated, and passed through a CNN predictor. Trained with tool knockout augmentation, the TBM learns to handle any arbitrary subset of tools at inference time. All tools are frozen; only the TBM is learned.


Results

TBF matches or outperforms strong trained baselines such as EfficientNet — while being interpretable and modular by design. Unlike black-box CNNs, TBF grounds its predictions in clinically-meaningful tools, and its decisions can be directly interrogated by inspecting which tools were selected and how much each contributes.

On-par or better performance

Matches or exceeds EfficientNet across all three tasks, with top accuracy on Camelyon17 and ISIC-MN.

Interpretable by design

Predictions are grounded in clinically-relevant tools. Tool importance can be measured via leave-one-out analysis.

Modular and extensible

New tools can be added to the toolbox without retraining. Tool knockout augmentation handles any subset of tools at inference.

MedGemma LLaVA-Med FT EfficientNet Y-Net TBF (ours)
50.0
66.2
88.6
88.2
92.3
Camelyon17
Accuracy (%)
44.4
51.5
78.4
65.8
77.0
ISIC-BM
AUC
48.7
58.0
91.2
86.6
91.8
ISIC-MN
AUC

Data Efficiency

A key advantage of TBF is its strong performance in low-data regimes. Because the model operates on clinically-grounded tool outputs rather than raw pixels, it requires far fewer labeled examples to learn effectively. With only 4 labeled examples, TBF already outperforms EfficientNet, and this advantage persists across all training set sizes, with TBF also exhibiting lower variance across random seeds.

0.95 0.85 0.75 0.65 Accuracy 4 12 32 64 160 300 Training set size TBF EfficientNet (a) Camelyon17 0.85 0.75 0.65 0.55 AUC 2 4 8 16 32 64 Training set size TBF EfficientNet (b) ISIC-MN
TBF vs. EfficientNet across training set sizes (log scale). Mean ± 95% CI over 8 seeds.

Tool Importance & Selection

The modular nature of TBM allows us to evaluate each tool's contribution via leave-one-out analysis during training. We can then analyze how this empirical importance aligns with the medical VLM's prior knowledge, reflected in its tool selection frequency. Tools where the two signals agree appear in the upper-right. Interestingly, nucleus contour is the most empirically important histopathology tool yet is selected less often than bbox and centroid, revealing a gap between learned and prior knowledge. In dermatology, pigment network and lesion mask score highly on both axes, suggesting strong alignment with clinical priors.

Camelyon17 — Histopathology
Tool Importance → VLM Selection Freq → 0 high 0 high contour bbox centroid type type prob
ISIC — Dermatology
Tool Importance → VLM Selection Freq → 0 high 0 high lesion mask pigment net brown marker neg. net malignancy streaks milia

Dot color reflects overall importance: dark = high, light = low.