CT4CMS: Preoperative Computed Tomography-Based Consensus Molecular Subtyping Prediction in Colorectal Cancer Using Interpretable Deep Learning

This study presents CT4CMS, an interpretable deep learning framework that utilizes preoperative CT scans to noninvasively predict colorectal cancer consensus molecular subtypes, enabling molecular stratification and guiding personalized therapeutic decisions before surgery.

The Gist

The Big Problem: Guessing the Enemy's Identity

Imagine you are a general preparing for a battle against an invading army (Colorectal Cancer). To win, you need to know exactly what kind of enemy you are facing. Are they fast runners? Heavy armor? Do they hide in the woods?

In the medical world, this "enemy identity" is called Consensus Molecular Subtyping (CMS). There are four main types of colorectal cancer (CMS1, 2, 3, and 4), and each behaves differently:

• Some are slow and easy to treat.
• Some are aggressive and dangerous.
• Some respond well to chemotherapy; others don't.

The Catch: Until now, to find out which type of enemy you are fighting, doctors had to wait until after the surgery to take a piece of the tumor, send it to a lab, and run expensive, time-consuming genetic tests (like reading a secret code). By the time the results come back, the surgery is already done, and the patient has already been treated. If the patient needed a different treatment plan before the surgery, the doctors were flying blind.

The Solution: The "X-Ray Detective" (CT4CMS)

This paper introduces a new AI tool called CT4CMS. Think of it as a super-smart detective that can look at a standard preoperative CT scan (a 3D X-ray of the abdomen) and instantly tell you exactly which type of cancer enemy you are facing—before the knife ever touches the patient.

Here is how the AI detective works, broken down into simple steps:

1. The "Schooling" Phase (Self-Supervised Learning)

Before the AI could diagnose cancer, it had to go to school. The researchers fed it thousands of CT scans of normal abdomens and other organs.

• The Analogy: Imagine teaching a child to recognize a cat by showing them thousands of pictures of cats, but covering up parts of the picture and asking, "What's missing?" The child learns the shape, fur texture, and ears by trying to fill in the blanks.
• The Tech: The AI used a technique called "Masked Image Modeling." It looked at CT scans, covered up random chunks, and tried to guess what was underneath. This taught the AI to understand the 3D structure of the human body and tumors without needing to know the cancer type yet.

2. The "Spotlight" Phase (Attention Mechanism)

Once the AI was smart enough, they showed it scans of known cancer cases (where the genetic type was already known).

• The Analogy: Now the AI is a detective looking at a crime scene. Instead of looking at the whole room, it puts on a high-tech spotlight. It zooms in on the specific spots in the tumor that give away the most clues.
• The Tech: This is called "Multi-Instance Learning." The AI breaks the 3D tumor into thousands of tiny 3D blocks. It assigns a "score" to each block. The blocks with the highest scores are the "clues" the AI used to make its decision. This makes the AI interpretable—doctors can see where the AI is looking, so they trust it more.

3. The Diagnosis

The AI looks at the preoperative CT scan, finds the tumor, shines its spotlight on the most important textures and shapes, and says: "This tumor is CMS4."

Why This Changes Everything

The paper proves that this AI is not just guessing; it is actually seeing the biology of the cancer through the X-ray.

• The "CMS4" Discovery: The AI found that patients with the CMS4 type (the most aggressive, "mesenchymal" type) usually have a very poor outlook if they only get surgery.
• The Chemotherapy Clue: However, the AI also discovered that CMS4 patients are the ones who benefit the most from chemotherapy.
• The Other Types: Patients with CMS1, 2, or 3 often don't get much help from chemotherapy and might just suffer the side effects for no reason.

The Real-World Impact:
Imagine a patient with Stage II or III cancer.

• Old Way: Doctor does surgery. Wait 2 weeks for genetic results. Then decide: "Oh, you have CMS4, you should have had chemo." (Too late).
• New Way (CT4CMS): Doctor looks at the CT scan. The AI says: "This is CMS4." The doctor immediately plans for surgery plus chemotherapy. The patient gets the right treatment before the surgery even happens.

The "Black Box" vs. The "Glass Box"

Many AI models are "Black Boxes"—they give an answer, but you don't know how they got there. This paper built a "Glass Box."

• By analyzing the specific spots the AI highlighted (the "high-attention regions"), the researchers found that the AI was looking at things that make sense biologically.
• For example, when the AI identified CMS1, it was looking at areas that looked "messy" and "uneven" on the scan, which matches the fact that CMS1 tumors are full of immune cells.
• When it identified CMS4, it looked at areas that were "heterogeneous" (mixed up), matching the fact that these tumors have lots of scar tissue and blood vessels.

Summary

This paper presents a non-invasive crystal ball. Using a routine CT scan and a clever AI that learned to "fill in the blanks," doctors can now predict the genetic personality of a colon cancer tumor before surgery. This allows them to tailor the treatment plan perfectly, sparing patients from unnecessary drugs and giving the most aggressive cases the heavy artillery they need right from the start.

Technical Summary

1. Problem Statement

Consensus Molecular Subtyping (CMS) is the gold standard for classifying colorectal cancer (CRC) into four distinct molecular subtypes (CMS1–CMS4) based on gene expression profiles. These subtypes have critical implications for prognosis and treatment selection (e.g., CMS4 is associated with poor prognosis but high sensitivity to chemotherapy).

However, current clinical implementation faces significant barriers:

• Invasiveness and Cost: Standard CMS determination requires RNA sequencing of surgical specimens, which is expensive, resource-intensive, and requires specialized bioinformatics.
• Timing Limitations: Histopathology-based methods rely on post-surgical specimens, rendering them useless for preoperative decision-making (e.g., neoadjuvant therapy planning).
• Sampling Bias: Biopsies may not represent the whole tumor's molecular landscape.
• Inoperable Cases: Patients with inoperable tumors cannot undergo surgical resection to obtain tissue for standard CMS testing.

The Gap: There is a lack of non-invasive, preoperative methods to predict CMS subtypes using routine imaging data that can guide personalized treatment strategies before surgery.

2. Methodology

The authors developed CT4CMS, an interpretable deep learning framework designed to predict CMS subtypes directly from preoperative contrast-enhanced CT scans.

A. Data Collection

• Total Cohort: 2,444 CRC patients from three independent institutions.
• Discovery Cohort (SYSU-CRC, n=416): Contains paired CT images and gene expression profiles (ground-truth CMS labels). Used for model training and cross-validation.
• Internal Validation (SYSU-SAH, n=1,671) & External Validation (Liaoning, n=357): Contain CT images only. Used to validate the model's generalizability and prognostic utility via survival analysis.
• Pretraining Data: Over 5,000 abdominal CT volumes from public datasets (AbdomenCT-1K, BTCV, MSD, WORD, TCIA-Colon) were used for self-supervised pretraining.

B. Model Architecture

The framework consists of four main stages:

1. Segmentation: Manual delineation of tumor regions by radiologists to define the Region of Interest (ROI).
2. Self-Supervised Pretraining (Feature Extraction):
   • Backbone: A 3D Swin-Transformer.
   • Strategy: Multiscale Masked Image Modeling (MIM). The model is pretrained on large-scale public CT datasets using a masked image modeling approach where random patches are masked, and the model learns to reconstruct them.
   • Multiscale Approach: Unlike single-scale methods, this strategy reconstructs masked patches at different resolutions (fine-scale for low-level features like edges; coarse-scale for high-level semantics like shape). This enables the encoder to learn robust, generalizable 3D representations without needing labeled data.
3. Classification (Multi-Instance Learning - MIL):
   • The pretrained backbone extracts 3D feature instances from the tumor volume.
   • An Attention-based MIL mechanism aggregates these instances into a global representation.
   • Interpretability: The attention mechanism assigns weights to specific 3D patches, highlighting which regions of the tumor contributed most to the prediction.
4. Prediction: A Multi-Layer Perceptron (MLP) classifies the aggregated features into one of the four CMS subtypes.

C. Interpretability Analysis

• Attention Mapping: High-attention regions (patches with attention scores > 0.5) were identified.
• Radiomic Correlation: 851 radiomic features (first-order, second-order, texture) were extracted from these high-attention regions. Statistical tests (Wilcoxon rank-sum, ANOVA) were used to correlate specific radiomic signatures with predicted CMS subtypes to validate biological plausibility.

3. Key Contributions

1. First Preoperative CT-Based CMS Prediction: Demonstrated the feasibility of inferring transcriptomic CMS subtypes directly from routine preoperative CT scans using deep learning.
2. Multiscale Self-Supervised Learning: Introduced a novel pretraining strategy using Multiscale MIM with a 3D Swin-Transformer, which outperformed standard supervised pretraining and handcrafted radiomics.
3. Interpretability via Attention: Developed an attention-based MIL framework that not only predicts subtypes but also visualizes the specific tumor regions driving the decision, linking them to distinct radiomic phenotypes.
4. Clinical Workflow Integration: Proposed a new clinical decision point where CT4CMS predictions can guide adjuvant chemotherapy decisions for Stage II/III patients before surgery.

4. Results

A. Classification Performance

• Discovery Cohort: CT4CMS achieved a cross-validation AUC of 0.867.
• Comparison: It significantly outperformed:
  • Handcrafted Radiomics + MLP (AUC = 0.586).
  • SwinUnetr (AUC = 0.736).
  • MaskedIM (AUC = 0.807).
• Per-Subtype Performance: Balanced performance across all subtypes (AUCs ranging from 0.857 to 0.863), with particularly high accuracy in identifying CMS4 (the most clinically critical subtype).

B. Prognostic Validation

• Survival Analysis: In all three cohorts (Discovery, Internal, External), patients predicted as CMS4 exhibited significantly worse Disease-Free Survival (DFS) compared to CMS1-3.
  • Discovery: P = 0.025
  • Internal: P = 3.69 × 10⁻⁶
  • External: P = 0.043
• Independence: Multivariable Cox regression confirmed that CT4CMS-predicted CMS4 status is an independent prognostic marker, even after adjusting for stage, age, and sex.

C. Therapeutic Stratification (Chemotherapy Benefit)

• CMS4 Sensitivity: Patients predicted as CMS4 who received adjuvant chemotherapy showed a significant improvement in DFS compared to those treated with surgery alone.
  • Discovery: P = 3.68 × 10⁻⁵
  • Validation Cohorts: Significant P-values observed in both SYSU-SAH and Liaoning cohorts.
• CMS1-3 Resistance: Patients with predicted CMS1, CMS2, or CMS3 subtypes showed no significant benefit from adjuvant chemotherapy (P > 0.05).
• Stage Specificity: This benefit was consistent in both Stage II and Stage III disease, offering a potential solution to the clinical ambiguity of treating Stage II patients.

D. Interpretability Findings

The analysis of high-attention regions revealed distinct radiomic signatures for each subtype, aligning with known biology:

• CMS1: Higher Grey Level Emphasis (GLE), suggesting immune cell infiltration.
• CMS2: Features related to skewness and percentiles, indicating epithelial differentiation.
• CMS3: Long Run Emphasis and larger diameters, reflecting mucus content.
• CMS4: High Dependency Entropy and Variance, indicating heterogeneity, stromal invasion, and angiogenesis.

5. Significance and Clinical Impact

• Non-Invasive Precision Medicine: CT4CMS enables molecular stratification without the need for invasive biopsies or expensive RNA sequencing, making precision oncology accessible in resource-limited settings.
• Preoperative Decision Making: By predicting CMS subtypes before surgery, clinicians can identify high-risk CMS4 patients who are likely to benefit from adjuvant chemotherapy, while sparing low-risk CMS1-3 patients from unnecessary toxicity.
• Generalizability: The model's success across two independent external validation cohorts demonstrates robustness against data distribution shifts.
• Future Directions: While the study is retrospective, it establishes a foundation for prospective trials. Future work aims to automate tumor segmentation to fully integrate the pipeline into clinical workflows.

In summary, CT4CMS represents a breakthrough in radiogenomics, successfully bridging the gap between routine CT imaging and complex molecular profiling to optimize treatment strategies for colorectal cancer patients.