Evaluating Multi-Task Morphological Concept Learning for Pulmonary Nodule Malignancy Assessment in 3D CT

Namitha Narayanan

cs.CVSep 29, 2026

Evaluating Multi-Task Morphological Concept Learning for Pulmonary Nodule Malignancy Assessment in 3D CT

Authors: Namitha Narayanan

Organizations: Independent Researcher, Kerala, India

Abstract

Morphological characteristics such as spiculation and lobulation play an important role in assessing pulmonary nodules on computed tomography (CT), particularly in relation to malignancy risk. This study examines whether learning radiologist-annotated morphological features together with malignancy risk from lesion-centred 3D CT volumes improves classification performance. The Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset was used, comprising 3,918 reader-level nodule annotations from 742 patients after excluding indeterminate malignancy ratings. Patient-level splitting was used for training, validation, and testing, with 112 patients and 628 reader annotations in the held-out test set. A single-task 3D convolutional neural network was compared with a multi-task model predicting malignancy risk, spiculation, and lobulation. The single-task model achieved a balanced accuracy of 0.548 and receiver operating characteristic area under the curve (ROC-AUC) of 0.552, while the multi-task model achieved 0.539 and 0.558, respectively. Patient-level bootstrap analysis showed an ROC-AUC difference of 0.005 (95% confidence interval (CI): -0.087 to 0.090) and a balanced-accuracy difference of -0.009 (95% CI: -0.067 to 0.043). The auxiliary tasks were strongly imbalanced and showed limited predictive performance. Overall, including morphological features did not clearly improve malignancy-risk classification, showing the importance of class balance, label formulation, and reader-level annotation structure in multi-task pulmonary CT analysis.

Figures & tables

Fig. 1: Example of lesion localisation and crop extraction from LIDC-IDRI CT data. The left panel shows the original CT slice with the radiologist-defined lesion contour and the corresponding crop region. The right panel shows the central slice of the preprocessed lesion-centred crop used as model input.

Panel of Figure 2 — Fig. 2: Overview of the single-task and multi-task 3D CNN models. Both models use the same 3D convolutional encoder, summarized in Table I , to obtain a 64-dimensional feature representation from each lesion-centred CT volume. The single-task model uses one prediction head for malignancy risk, whereas the multi-task model uses separate heads for malignancy risk, spiculation, and lobulation.

Stage	Operation	Output
Input	Lesion-centred CT crop	$1\times 64\times 64\times 64$
Conv 1	$3\times 3\times 3$ , 16 channels + ReLU	$16\times 64\times 64\times 64$
Pool 1	$2\times 2\times 2$ max pooling	$16\times 32\times 32\times 32$
Conv 2	$3\times 3\times 3$ , 32 channels + ReLU	$32\times 32\times 32\times 32$
Pool 2	$2\times 2\times 2$ max pooling	$32\times 16\times 16\times 16$
Conv 3	$3\times 3\times 3$ , 64 channels + ReLU	$64\times 16\times 16\times 16$

TABLE I: Common 3D Encoder Architecture

Split	Patients	Annotations	Negative	Positive
Training	519	2754	1754	1000
Validation	111	536	294	242
Test	112	628	375	253
Total	742	3918	2423	1495

TABLE II: Cohort Characteristics and Malignancy-Risk Distribution

Metric	Spiculation	Lobulation
Valid records	584	557
Positive records	67	50
Threshold	0.503	0.523
ROC-AUC	0.546	0.480
Average Precision	0.137	0.088
Balanced Accuracy	0.499	0.464

TABLE III: Auxiliary Morphological Concept Prediction Performance

Metric	Single-Task	Multi-Task	Difference
ROC-AUC	0.552	0.558	+0.005
Average Precision	0.431	0.443	+0.012
Balanced Accuracy	0.548	0.539	-0.009
Accuracy	0.490	0.484	-0.006
Sensitivity	0.842	0.822	-0.020
Specificity	0.253	0.256	+0.003

TABLE IV: Comparison of Malignancy-Risk Classification Performance

Panel of Figure 7 — Fig. 3: Comparison of single-task and multi-task malignancy-risk classification on the held-out test set. The upper panel shows the receiver operating characteristic (ROC) curves, and the lower panel shows the precision–recall curves. AP denotes average precision.

Slice	Patients	Records	Single-Task	Multi-Task
			AUC / BA	AUC / BA
$\leq 2$ mm	68	366	0.555 / 0.555	0.536 / 0.539
$>2$ mm	44	262	0.539 / 0.531	0.586 / 0.519

TABLE V: Performance by CT Slice-Thickness Subgroup

Explore similar work

Distilling CT Foundation Models into Editable Concept Bottlenecks for Lung Nodule Malignancy Prediction
Aug 8, 2026Fakrul Islam Tushar, Stephen Adamo, Geoffrey D. RubinLung-Nodule Artificial Intelligence Concept Bottleneck Models
Department of Radiology and Imaging Sciences, University of Arizona, Tucson, AZ, USA
M3Net: A Macro-to-Meso-to-Micro Clinical-inspired Hierarchical 3D Network for Pulmonary Nodule Classification
May 12, 2026Jinyue Li, Yuzhou Yu, Jingjing Yang +7Lung-Nodule Artificial Intelligence Multiscale
Hefei Cancer Hospital of CAS, Institute of Health and Medical Technology, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, 230031, PR China · University of Science and Technology of China, Hefei, 230026, PR China · Graduate School, Bengbu Medical College, Bengbu, 233030, PR China +4
Foundation Models vs. Radiomics for Lung Computed Tomography: A Benchmark of Feature Extractors, Classification Heads, and Segmentation Choices
Jul 1, 2026Nils Neukirch, Martin Maurer, Nils StrodthoffRadiomics Cohort
Division AI4Health, Carl von Ossietzky Universität Oldenburg, Ammerländer Heerstraße 114-118, Oldenburg 26129, Niedersachsen, Germany · University Institute for Diagnostic and Interventional Radiology, Klinikum Oldenburg AöR, Rahel-Straus-Straße 10, Oldenburg 26133, Niedersachsen, Germany

Aug 8, 2026, cs.CVJ/K move · Enter open · S save

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

Fakrul Islam Tushar, Stephen Adamo, Geoffrey D. Rubin

Department of Radiology and Imaging Sciences, University of Arizona, Tucson, AZ, USA

Lung-Nodule Artificial Intelligence Concept Bottleneck Models Bottlenecks

Open paper AskPDF arXiv

Foundation models provide transferable CT representations, but predictions based directly on these embeddings are difficult to interpret. We developed concept bottleneck models that map two frozen CT foundation-model representations to eight radiologist-defined pulmonary-nodule attributes and predict malignancy from the estimated concepts and nodule size. The models included CT-FM, a whole-CT self-supervised encoder using a 96^3-voxel nodule-centered patch, and FMCIB, a nodule-focused contrastive encoder using a 50-mm crop. Eight ridge-regression concept heads were trained on 2,610 LIDC-IDRI nodules. Malignancy models were trained on LUNA25 and evaluated on a held-out internal test set and the external DLCS cohort. Concept fidelity was assessed using five-fold cross-validated R^2, and malignancy discrimination was assessed using AUROC with 95% confidence intervals estimated by patient-grouped bootstrap resampling. Concept fidelity was modest but higher for FMCIB than CT-FM for subtlety (R2, 0.24 vs. 0.11), spiculation (0.17 vs. 0.08), texture (0.17 vs. 0.07), and lobulation (0.15 vs. 0.05). Internally, the CT-FM and FMCIB concept+size models achieved AUROCs of 0.86 (95% CI, 0.80-0.92) and 0.86 (0.79-0.92), respectively. Externally, AUROCs were 0.72 (0.68-0.75) and 0.73 (0.70-0.76), compared with 0.73 for nodule size alone and 0.60 and 0.67 for the corresponding embedding only probes. Additive predictions could be decomposed into feature-level contributions and modified through controlled concept interventions. Concept bottlenecks provided transparent malignancy predictions with discrimination similar to nodule size alone, while differences in concept fidelity suggest that concept recovery depends on the underlying foundation-model representation.

Abstract

Explore similar work

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

M3Net: A Macro-to-Meso-to-Micro Clinical-inspired Hierarchical 3D Network for Pulmonary Nodule Classification

Foundation Models vs. Radiomics for Lung Computed Tomography: A Benchmark of Feature Extractors, Classification Heads, and Segmentation Choices