🤖 machine learning

Bridging visual saliency and large language models for explainable deep learning in medical imaging

This paper proposes a multimodal explainability framework that integrates CNN-based tumor classification and segmentation with visual saliency mapping and large language models to generate human-interpretable, radiological-style diagnostic reports for brain tumor MRI analysis, thereby enhancing the transparency and clinical adoption of AI in medical imaging.

Original authors: Paul Valery Nguezet, Elie Tagne Fute, Yusuf Brima, Benoit Martin Azanguezet, Marcellin Atemkeng

Published 2026-05-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Paul Valery Nguezet, Elie Tagne Fute, Yusuf Brima, Benoit Martin Azanguezet, Marcellin Atemkeng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a brilliant but silent radiologist. This doctor can look at a brain scan and instantly spot a tumor with incredible accuracy. However, if you ask, "Why do you think that's a tumor?" or "Which part of the brain is affected?", the doctor just points at the screen and says, "I'm sure, trust me." They can't explain their reasoning in words. This is the problem with many current AI models in medicine: they are "black boxes" that give answers without explanations.

This paper introduces a new system designed to give that silent AI a voice and a map. It bridges the gap between a computer's raw calculation and a human doctor's understanding. Here is how the system works, step-by-step, using simple analogies:

1. The "Dual-Brain" Training (The Foundation)

First, the researchers taught nine different AI models (called CNNs) to do two jobs at once. Imagine a student studying for an exam who is also learning to draw a map of the classroom.

Job A: Identify the type of tumor (like Glioma, Meningioma, or Pituitary).
Job B: Draw a rough outline of where the tumor is.

By forcing the AI to learn where the tumor is while it learns what it is, the model becomes smarter and more precise. The best student in this class was a model called InceptionResNetV2.

2. The "Spotlight" (Visual Saliency)

Once the AI is trained, it still can't speak. So, the researchers used a "spotlight" technique. They asked the AI, "Which pixels in this image made you decide it's a tumor?"

The AI shines a glowing heat map over the image, highlighting the areas it looked at most closely.
The researchers tested three different types of spotlights (Grad-CAM, Grad-CAM++, and ScoreCAM).
The Winner: The Grad-CAM++ spotlight was the sharpest. It didn't just glow vaguely; it focused tightly on the actual tumor tissue, ignoring the healthy brain around it.

3. Turning Light into a Shape (Segmentation)

A glowing heat map is still a bit fuzzy. It's like seeing a shadow on the wall; you know something is there, but you don't know the exact shape.

The system takes that fuzzy glow and uses a smart filter to cut it out, turning the "glow" into a crisp, black-and-white silhouette (a mask) of the tumor.
This step is crucial because it turns a vague idea into a concrete shape that can be measured.

4. The "Anatomical Translator" (Atlas Mapping)

Now the system has a shape, but it doesn't know what that shape is touching. Is it near the memory center? The speech center?

The researchers took the tumor shape and overlaid it onto a giant, detailed 3D map of the human brain (called the Harvard-Oxford Atlas).
Think of this like placing a sticker of a tumor onto a map of a city. The system can now say, "This tumor is sitting right on top of the 'Insular Cortex' and the 'Anterior Cingulate Gyrus'."
It calculates exactly what percentage of those brain areas are covered by the tumor.

5. The "Medical Reporter" (Large Language Models)

Finally, the system needs to speak. It takes all the data—the tumor type, the confidence score, the exact brain areas involved, and the size of the tumor—and feeds it into a "translator" (a Large Language Model or LLM).

The researchers tested three different translators: Grok3, Mistral, and LLaMA.
These models act like a professional medical scribe. They take the raw data and write a coherent, easy-to-read report that sounds like a doctor talking to another doctor.
The Results:
- Grok3 was the most "wordy" and diverse, using a rich vocabulary.
- LLaMA was the easiest to read, using simple, clear sentences.
- Both managed to create reports that sounded professional and logical.

The Final Product

The result is a complete package. Instead of just getting a label that says "Meningioma," a doctor gets a report that says:

"The AI has identified a Meningioma. It is located primarily on the Insular Cortex and Anterior Cingulate Gyrus, covering about 64% of the latter. Based on this location, the tumor may affect emotional regulation and decision-making."

Why This Matters (According to the Paper)

The paper claims this system makes AI "explainable." It doesn't just guess; it shows its work. It connects the computer's math to human anatomy and then translates that into human language. This helps doctors trust the AI because they can see why the AI made its decision and where the problem is, rather than just taking a blind guess.

Important Note: The paper focuses entirely on creating this explanation framework and testing how well the models perform. It does not claim that this system is currently being used in hospitals to treat patients, nor does it predict future clinical outcomes. It is a proof-of-concept that shows how to make AI transparent and readable.

Technical Summary: Bridging Visual Saliency and Large Language Models for Explainable Deep Learning in Medical Imaging

Problem Statement
The clinical adoption of deep learning (DL) models in medical imaging is hindered by their "opaque" nature. While Convolutional Neural Networks (CNNs) demonstrate high accuracy in tasks like brain tumor classification, they lack transparency, failing to provide the justification and interpretability required by medical professionals. Existing explainability techniques, such as visual saliency maps (e.g., Grad-CAM), offer spatial localization but suffer from significant limitations: they are often coarse, lack semantic grounding, and fail to articulate specific radiological features or anatomical contexts. Furthermore, purely visual explanations do not bridge the gap to actionable clinical reasoning or narrative reporting.

Methodology
The authors propose a novel, three-stage multimodal explainability framework that integrates computational visual saliency, domain-aware anatomical contextualization, and natural language generation. The pipeline operates as follows:

Dual-Output Hybrid CNN Training:
- Nine established CNN architectures (including DenseNet, VGG, ResNet, Inception, and EfficientNet variants) were extended with a dual-output formulation.
- Each model shares an encoder that feeds two independent heads: a classification head (softmax over three tumor classes: Glioma, Meningioma, Pituitary) and a segmentation head (producing a spatial probability map).
- Training utilizes a combined loss function: categorical cross-entropy for classification and Dice Similarity Coefficient (DSC) loss for segmentation. This joint optimization strengthens the feature representations, anchoring the subsequent explainability pipeline.
- The dataset consists of 4,834 contrast-enhanced T1-weighted MRI images.
Visual Saliency and Adaptive Segmentation:
- To ensure interpretability, saliency computations (Grad-CAM, Grad-CAM++, and ScoreCAM) are explicitly routed through the classification head only, preventing gradient contamination from the segmentation branch.
- The resulting coarse heatmaps are refined into precise binary tumor masks via an adaptive percentile thresholding pipeline. Unlike fixed thresholds, the optimal percentile ( $\alpha^*$ ) is determined per sample by maximizing the DSC against the ground truth within a search range of [70, 97].
- Post-processing includes the removal of small connected components (<50 pixels) and morphological closure to ensure coherent tumor regions.
Anatomical Mapping and Narrative Generation:
- The binary masks are overlaid onto the Harvard–Oxford cortical atlas. This step translates pixel-level coordinates into named neuroanatomical structures (e.g., "Insular Cortex," "Cingulate Gyrus") and calculates the percentage of tumor overlap within each region.
- These findings are encoded into a structured JSON file containing the predicted class, confidence, saliency method, and anatomical details.
- Three Large Language Models (LLMs)—Grok3, Mistral, and LLaMA—are conditioned on this structured JSON to generate coherent, radiological-style diagnostic reports. The structured input is designed to minimize hallucinations and ensure factual consistency.

Key Contributions

Unified Multimodal Framework: The paper introduces a pipeline that tightly couples classification, saliency-derived segmentation, anatomical mapping, and natural language generation, moving beyond isolated visual explanations.
Saliency-Derived Segmentation: The authors propose using saliency heatmaps as attention guides for segmentation without requiring additional annotated segmentation training data. This offers a computationally lightweight and inherently explainable alternative to dedicated segmentation networks (like UNet) in settings where labeled masks are scarce.
Anatomical Contextualization: By mapping segmentation outputs to the Harvard–Oxford atlas, the framework provides causally linked, domain-aware explanations that align with clinical reasoning, translating raw pixel data into clinically meaningful labels.
LLM-Driven Reporting: The system demonstrates the feasibility of using LLMs to synthesize quantitative imaging features and anatomical data into professional medical narratives, evaluated on lexical diversity, readability, and coherence.

Results

Classification Performance: Among the nine architectures tested, InceptionResNetV2 achieved the highest performance, with an accuracy of 0.96 and balanced precision, recall, and F1-scores of 0.95 across all three tumor classes.
Segmentation Quality: Grad-CAM++ yielded the best segmentation overlap, achieving a mean DSC of 0.384 and IoU of 0.259, outperforming Grad-CAM and ScoreCAM. This is attributed to its per-neuron weighting scheme which reduces diffuse background activation.
Language Model Evaluation:
- Grok3 demonstrated the highest lexical diversity (TTR: 0.467) and coherence.
- LLaMA achieved the highest readability score (Flesch Reading Ease: 18.85).
- All models successfully generated reports that integrated tumor type, anatomical involvement, and potential clinical impacts.

Significance and Claims
The paper claims that by integrating visual, anatomical, and linguistic modalities, the proposed framework produces explanations that are both technically grounded and meaningfully interpretable. The authors assert that this approach advances the transparency and clinical accountability of AI-assisted brain tumor diagnosis. By aligning model outputs with the cognitive workflows and informational expectations of medical professionals, the framework aims to foster greater trust and facilitate smoother integration into clinical diagnostic pathways. The work emphasizes that saliency-based segmentation, while potentially less precise than dedicated segmentation models trained on masks, offers distinct advantages in annotation efficiency and interpretability, making it a practical alternative for clinically constrained environments.