💻 computer science

Distilling CT Foundation Models into Editable Concept Bottlenecks for Lung Nodule Malignancy Prediction

This paper introduces editable concept bottleneck models that distill frozen CT foundation model representations into radiologist-defined attributes to provide transparent lung nodule malignancy predictions with discrimination performance comparable to nodule size alone, while demonstrating that concept recovery fidelity depends on the underlying foundation model architecture.

Original authors: Fakrul Islam Tushar, Stephen Adamo, Geoffrey D. Rubin

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fakrul Islam Tushar, Stephen Adamo, Geoffrey D. Rubin

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 you are trying to teach a super-smart robot how to spot a dangerous thief in a crowded city. You give the robot a pair of magical glasses that can see everything in the city at once, but the robot just sees a giant, confusing blur of colors and shapes. It can tell you, "That person is suspicious!" but it can't tell you why. Is it because they are wearing a mask? Because they are running? Because they look nervous? In the world of medical science, specifically looking at CT scans of lungs, doctors face this exact problem. They have powerful AI "foundation models" that can look at lung images and guess if a small lump (a nodule) is cancer, but these models act like a black box: they give an answer without showing their work. This is risky because doctors need to understand the reasoning to trust the diagnosis. To fix this, scientists are trying to build "Concept Bottleneck" models. Think of this as forcing the robot to stop and describe the thief's features in plain English—like "wearing a hat," "holding a bag," or "limping"—before it makes its final guess. If the robot can explain its reasoning using these human-friendly clues, doctors can trust it more, and they can even tweak the clues to see how the answer changes.

This paper takes that idea and applies it to lung cancer screening. The researchers wanted to see if they could take two different types of "magical glasses" (AI foundation models) that were already trained to look at CT scans, and force them to describe lung nodules using eight specific features that real radiologists use, such as how "spiky" the edges are or how "round" the lump is. They then used these descriptions, along with the size of the lump, to predict if the nodule was malignant (cancerous).

Here is what they found:

First, they tried to teach the AI to describe the lung nodules. They used two different AI models: one that looked at the whole chest scan (CT-FM) and one that zoomed in tightly on just the nodule (FMCIB). They asked the AI to predict eight features, like "spiculation" (spiky edges) and "subtlety" (how hard it is to see). The results were a bit mixed. The zoomed-in model (FMCIB) was better at describing the features than the whole-chest model. For example, when guessing how "spiky" a nodule was, the zoomed-in model got a score of 0.17, while the whole-chest model only got 0.08. However, even the better model wasn't perfect; it still struggled to accurately describe features like "calcification" (hard spots) or "sphericity" (how round it is). This suggests that while the AI can learn to describe some features, it depends heavily on how the AI was originally trained.

Next, they tested if these "descriptions" actually helped predict cancer. They built a system that took the AI's descriptions plus the actual size of the nodule to make a prediction. The results were surprising. The system that used the AI's descriptions performed just as well as a system that only looked at the size of the nodule. On their internal test, both systems got a score of 0.86 (a measure of how good the prediction is). On an external test with different patients, both got around 0.72 to 0.73. Interestingly, the AI models that tried to predict cancer without using the descriptions (just looking at the raw data) actually did worse, scoring only 0.60 to 0.67.

So, what does this mean? The main takeaway is that the AI's "descriptions" didn't make the prediction more accurate than just measuring the size of the lump. In fact, the size of the nodule was the most powerful predictor of all. However, the "descriptions" did something very valuable: they made the AI transparent. Because the model was forced to use human-defined concepts, the doctors could see exactly which features contributed to the cancer risk. They could even play a game of "what if": they could tell the model, "Imagine this nodule is less spiky," and watch the predicted risk go down. This proves that while the AI didn't become a better predictor by using these concepts, it became a much better explainer. The study suggests that while we can distill complex AI into understandable concepts, the quality of those concepts depends on the underlying AI, and for now, the size of the lump remains the most important clue for spotting lung cancer.

Technical Summary: Distilling CT Foundation Models into Editable Concept Bottlenecks for Lung Nodule Malignancy Prediction

Problem Statement
While CT foundation models (FMs) offer transferable representations for lung imaging, classifiers operating directly on their embeddings function as "black boxes," lacking interpretability. Existing interpretability methods, such as post-hoc feature attribution (e.g., SHAP) or additive glass-box models, often fail to provide controlled, feature-level explanations. Concept Bottleneck Models (CBMs) address this by routing predictions through human-defined intermediate attributes. However, the clinical utility of CBMs depends on the fidelity with which the model recovers intended radiologic characteristics from the underlying FM embeddings. It remains unclear how different foundation model architectures and pretraining objectives affect this concept recovery and subsequent malignancy prediction performance.

Methodology
The study proposes a framework to distill two distinct, frozen CT foundation models into editable concept bottlenecks for predicting lung nodule malignancy.

Foundation Models: Two frozen encoders were utilized without fine-tuning:
1. CT-FM: A whole-CT self-supervised encoder processing a 963-voxel nodule-centered patch (512 dimensions).
2. FMCIB: A nodule-focused contrastive encoder processing a tight 50×50×50 mm crop at 1-mm isotropic resolution (4,096 dimensions).
Concept Mapping: Eight ridge-regression heads were trained on 2,610 nodules from the LIDC-IDRI dataset to map the frozen embeddings to eight radiologist-defined semantic attributes: subtlety, internal structure, calcification, sphericity, margin, lobulation, spiculation, and texture.
Prediction Model: An additive "glass-box" classifier predicted malignancy using the estimated concepts and nodule size ( $\hat{y} = f(c, \text{size})$ ).
Datasets and Evaluation:
- Training: Concept heads used LIDC-IDRI; malignancy models used LUNA25 (training/validation).
- Internal Testing: Held-out LUNA25 test set (581 nodules).
- External Testing: DLCS cohort (2,487 nodules).
- Baselines: Performance was benchmarked against nodule size alone and embedding-only linear probes (black-box comparators).
Metrics: Concept fidelity was measured via five-fold cross-validated $R^2$ . Malignancy discrimination was assessed using AUROC with 95% confidence intervals derived from patient-grouped bootstrap resampling. Editability was quantified by the change in predicted risk when concepts were shifted from the 10th to the 90th percentile.

Key Results

Concept Fidelity: Recovery of radiologist-defined concepts was modest overall but varied significantly between foundation models. FMCIB demonstrated higher fidelity than CT-FM for key morphologic attributes: subtlety ( $R^2$ 0.24 vs. 0.11), spiculation (0.17 vs. 0.08), texture (0.17 vs. 0.07), and lobulation (0.15 vs. 0.05). Fidelity was near zero for calcification, sphericity, and internal structure for both models. The study noted that in-sample $R^2$ substantially overestimated fidelity (reaching ~0.98 for FMCIB), reinforcing the necessity of cross-validated evaluation.
Malignancy Discrimination:
- Internal (LUNA25): Both CBMs achieved an AUROC of 0.86 (95% CI: 0.80–0.92 for CT-FM; 0.79–0.92 for FMCIB).
- External (DLCS): CT-FM and FMCIB achieved AUROCs of 0.72 (0.68–0.75) and 0.73 (0.70–0.76), respectively.
- Comparison: The CBMs performed similarly to nodule size alone (AUROC 0.86 internal, 0.73 external) and outperformed their corresponding embedding-only probes (e.g., 0.60 and 0.67 for probes on DLCS). This indicates that nodule size was the primary driver of discrimination, while the concept bottleneck added interpretability rather than significant predictive gain.
Editability and Decomposition: The additive nature of the models allowed for transparent decomposition of predictions. Controlled interventions (shifting concepts from low to high values) showed that lobulation, subtlety, and spiculation produced the largest changes in predicted risk, while sphericity, margin, and calcification had minimal impact.

Significance and Claims
The paper claims that distilling frozen CT foundation models into radiologist-defined concept bottlenecks yields malignancy predictors that are both interpretable and editable while maintaining performance comparable to nodule size alone. The primary contribution is the demonstration that such models can provide transparent, feature-level explanations and support controlled concept interventions, even if the concepts themselves contribute less predictive information than nodule size.

The authors modestly conclude that the concept bottleneck primarily serves to add transparency rather than improve accuracy. They further highlight that concept fidelity is dependent on the underlying foundation model representation, with nodule-focused encoders (FMCIB) capturing morphologic attributes more effectively than whole-CT encoders (CT-FM) for specific attributes, though architectural differences prevent attributing this solely to the field of view. The study acknowledges limitations including reliance on a single concept annotation cohort, the use of frozen linear heads, and the evaluation of only two foundation models.

Technical Summary: Distilling CT Foundation Models into Editable Concept Bottlenecks for Lung Nodule Malignancy Prediction

More like this