⚡ electrical engineering

An Interpretable Deep Learning Framework for Discovery and Clinical Validation of Deep Radiomic Signatures in Tumor Classification

This paper proposes an interpretable deep learning framework that integrates segmentation, Grad-CAM-guided signature discovery, and SHAP-based analysis to extract and validate reproducible quantitative imaging biomarkers, demonstrating improved tumor classification performance and biological insight across multiple medical imaging datasets compared to conventional whole-tumor radiomics.

Original authors: Chengkun Sun, Jinqian Pan, Renjie Liang, Zhengkang Fan, Xin Miao, Yi Guo, Mei Liu, Muxuan Liang, Russell Terry, Jie Xu

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chengkun Sun, Jinqian Pan, Renjie Liang, Zhengkang Fan, Xin Miao, Yi Guo, Mei Liu, Muxuan Liang, Russell Terry, Jie Xu

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

The Big Problem: The "Black Box" Doctor

Imagine a highly skilled detective (a Deep Learning AI) who is amazing at solving crimes (diagnosing tumors). This detective can look at a blurry photo of a crime scene (a medical scan) and instantly tell you if it's a "good guy" (benign tumor) or a "bad guy" (malignant tumor) with incredible accuracy.

However, there is a catch: The detective won't tell you why they made that decision. They just point to the photo and say, "It's a bad guy." Because they can't explain their reasoning, real-world doctors (the judges) are hesitant to trust them. They need to know which part of the photo proved the tumor was dangerous.

The Solution: A New "Highlighter" Framework

The authors of this paper built a new system to fix this. Think of it as a three-step process that turns the detective's "gut feeling" into a clear, explainable report.

Step 1: The Outline (Segmentation)

First, the system uses a tool to draw a rough outline around the entire tumor, like a child tracing a picture. This tells the computer, "Okay, the trouble is somewhere inside this circle."

Paper Claim: They use advanced AI models (nnU-Net and UNETR) to draw these outlines accurately on CT scans and ultrasounds.

Step 2: The Spotlight (Grad-CAM)

Next, the system asks the detective, "Okay, you know the trouble is in this circle, but exactly where inside the circle is the smoking gun?"
The system uses a technique called Grad-CAM to shine a "spotlight" on the specific spots inside the tumor that look most suspicious. It's like using a flashlight in a dark room to find the specific clue that matters, ignoring the rest of the clutter.

Paper Claim: This creates a "saliency map" (a heat map) showing which pixels the AI is paying attention to.

Step 3: The Filter (The "Signature")

Here is the clever part. The system compares the "Outline" (Step 1) and the "Spotlight" (Step 2). It asks: "Which parts of the outline are also lit up by the spotlight?"
It creates a new, refined shape called a "Signature." This signature isn't the whole tumor; it's just the specific, high-value parts of the tumor that the AI thinks are most important for making the diagnosis.

Paper Claim: They use a math trick called "Mutual Information" to automatically find the perfect cutoff point for this spotlight, ensuring they don't miss anything important or include too much junk.

The Proof: Does it Work?

The researchers tested this new "Signature" system on three different types of medical puzzles:

Brain Tumors (MRI): Distinguishing between two types of brain growths.
Breast Tumors (Ultrasound): Telling if a lump is harmless or dangerous.
Kidney Tumors (CT Scans): Doing the same for kidney lumps.

The Results:

Better Accuracy: When they gave the AI the "Signature" (the refined spotlight area) along with the original image, the AI got better at guessing correctly than when it just looked at the whole image or the whole outline.
The "Intersection" Trick: For brain and breast tumors, the system worked best when it only looked at the parts where the "Outline" and the "Spotlight" overlapped. It was like saying, "Ignore the edges of the tumor; only look at the weird, suspicious center."
The "Union" Trick: For kidney tumors, the system worked best when it kept both the outline and the spotlight. It was like saying, "Look at the whole tumor, but pay extra attention to the suspicious spots."

Making Sense of the "Why" (Interpretability)

The best part of this framework is that it doesn't just give a score; it explains the biological reason.

The researchers took the "Signature" areas and measured them using standard math tools (Radiomics).
They found that the AI was looking for specific things, like:
- Rough Edges: Malignant tumors often have jagged, messy borders (like a crumpled piece of paper), while benign ones are smooth (like a polished marble).
- Patchy Colors: Bad tumors often have a mix of light and dark spots (heterogeneity), while good ones look uniform.
- Shape: Bad tumors might be stretched out or irregular, while good ones are round.

By isolating these specific "Signature" areas, the system could say, "I called this tumor dangerous because the center of it has a rough, patchy texture," which is a reason a human doctor can understand and trust.

The Limitations (The Fine Print)

The authors are honest about where their system might stumble:

Garbage In, Garbage Out: If the initial outline (Step 1) is drawn wrong, the whole system fails.
Missing Data: The system needs the image, the outline, and the spotlight all at once. If a patient's file is missing one of these, the system can't help.
Not a Magic Crystal Ball: The "Spotlight" shows where the AI is looking, but it doesn't prove that the spot is biologically the cause of the disease. It's a strong hint, not a legal verdict.

Summary

This paper presents a new way to make AI doctors more trustworthy. Instead of just giving a final grade, the system draws a map, shines a flashlight on the most important clues, and creates a "Signature" of the tumor that explains why the diagnosis was made. It turns a mysterious "black box" into a transparent, explainable tool that doctors can actually use.

Technical Summary: An Interpretable Deep Learning Framework for Discovery and Clinical Validation of Deep Radiomic Signatures in Tumor Classification

Problem Statement

While deep learning has demonstrated significant potential in medical image analysis, its limited interpretability remains a primary barrier to clinical adoption. Existing deep learning approaches often achieve high predictive performance but function as "black boxes," providing little biological insight into the specific imaging signatures driving classification decisions. Conversely, traditional radiomics relies on handcrafted features extracted from manually or automatically segmented whole tumors. While these traditional signatures are interpretable, they often assume the entire tumor contributes uniformly to diagnosis, potentially overlooking spatially localized, clinically critical subregions (e.g., infiltrative margins, necrotic cores, or heterogeneous transition zones). Furthermore, traditional workflows are labor-intensive and rely heavily on expert-driven preprocessing. There is a fundamental gap between traditional signatures, which are interpretable but may miss localized biological complexity, and deep learning signatures, which are highly predictive but insufficiently explainable.

Methodology

The authors propose a unified framework that integrates deep learning-based segmentation, explainable classification, and radiomic analysis to discover interpretable imaging signatures. The workflow consists of four main stages:

Robust Segmentation: The framework first employs deep learning models (specifically nnU-Net and UNETR) to accurately delineate tumor regions. To enhance robustness, a union of predictions from both models is computed to define the tumor mask.
Signature Discovery via Grad-CAM: A Gradient-weighted Class Activation Mapping (Grad-CAM) guided pipeline identifies diagnostically important regions. To focus attention on the tumor, an exponential distance decay is applied to the Class Activation Map (CAM) based on the distance from the segmentation mask centroid.
Adaptive Thresholding: A mutual information-based adaptive thresholding strategy is employed to extract patient-specific signatures. The system exhaustively evaluates candidate thresholds (0–255) within the segmentation mask region. The optimal threshold is selected as the one that maximizes the mutual information between the binarized CAM and the ground-truth mask. This process yields a "Signature" region representing the saliency activation with the highest informational overlap with the tumor mask.
Validation and Interpretation:
- Deep Learning Validation: The extracted signatures are validated using downstream deep learning classification models (EfficientNet, ResNet50, DenseNet121, and Vision Transformer) across three input settings: Image-only, Image + Segmentation Mask, and Image + Mask + Signature.
- Radiomic Analysis: Radiomic features are extracted from the signature-defined regions (and compared against the full mask). These features are analyzed using traditional machine learning models (Random Forest, XGBoost, LinearSVM, Logistic Regression).
- Interpretability: SHAP (SHapley Additive exPlanations) values are used to identify the most discriminative biomarkers and link them to pathological characteristics.

Key Contributions

Novel Framework: The paper introduces a signature construction framework that jointly utilizes segmentation masks and Grad-CAM to generate salient maps defining biologically meaningful signature regions.
Adaptive Extraction: It proposes a mutual information-based adaptive thresholding strategy to extract patient-specific signatures, moving beyond fixed or arbitrary cutoffs.
Dual-Validation Approach: The framework validates the signatures through both deep learning classification (demonstrating discriminative effectiveness) and traditional machine learning with radiomic features (providing biological interpretability).
Comprehensive Evaluation: Extensive experiments were conducted on four datasets: BUSI (breast ultrasound), BraTS (brain tumor MRI), KiTS (renal CT), and a private UF Health renal CT cohort, covering different imaging modalities and tumor types.

Results

Deep Learning Performance: Across all datasets and backbone architectures, incorporating the proposed signature consistently improved classification performance compared to using images alone or images with segmentation masks.
- On the BraTS dataset, adding the signature to the mask increased AUCs (e.g., DenseNet improved from 0.951 to 0.960).
- On the BUSI dataset, the signature provided significant gains, with DenseNet AUC increasing from 0.856 to 0.883 and ViT F1 scores improving substantially.
- On the UF Health renal CT dataset, improvements were more modest but generally positive, suggesting that while global morphology is critical for renal tumors, the signature still offers complementary information.
Machine Learning Performance: In radiomic analysis, the "Mask $\cap$ Signature" strategy (intersection of mask and signature) often achieved the highest AUC, particularly for brain and breast tumors. This suggests the signature effectively refines the tumor mask by excluding non-critical regions while preserving discriminative areas. For renal tumors, the union of mask and signature sometimes performed best, indicating that complete tumor morphology remains highly informative.
Interpretability: SHAP analysis revealed distinct feature importance patterns. For brain tumors, signature-guided regions highlighted infiltrative margins and heterogeneous vascular components. For breast tumors, features related to surface area and texture heterogeneity were key. For renal tumors, features capturing large low-intensity regions and elongation were prominent. These findings linked the extracted signatures to specific pathological characteristics (e.g., infiltrative growth in gliomas vs. homogeneous enhancement in meningiomas).

Significance and Claims

The authors claim that this framework offers an interpretable and reproducible solution for non-invasive tumor characterization and imaging biomarker discovery. By converting deep learning attention into reproducible quantitative imaging biomarkers, the approach bridges the gap between the predictive power of deep learning and the biological transparency required for clinical trust. The study demonstrates that signature-based approaches can achieve improved discriminative performance over conventional whole-tumor radiomics while providing greater biological interpretability.

The authors acknowledge limitations, including dependence on the accuracy of underlying segmentation models, potential variability in the adaptive thresholding strategy, and the need for larger-scale datasets to further enhance generalizability. They also note that Grad-CAM reflects model attention rather than true causal pathology, requiring cautious clinical interpretation.