From Slices to Volumes: Comparing 2D, 2.5D, and 3D Residual Networks for Pulmonary Nodule Malignancy-Risk Classification

Bahodir Nazarov; Zineb Gharib; Fatima Zazoul

cs.LGOct 25, 2026

From Slices to Volumes: Comparing 2D, 2.5D, and 3D Residual Networks for Pulmonary Nodule Malignancy-Risk Classification

Authors: Bahodir Nazarov, Zineb Gharib, Fatima Zazoul

Organizations: Sichuan University, Chengdu, Sichuan, China

Abstract

Purpose: To compare 2D, 2.5D, and 3D residual networks for pulmonary nodule malignancy-risk classification and evaluate momentum-contrast pretraining, mild augmentation, and exclusion of indeterminate ratings. Methods: A retrospectively prepared LIDC-IDRI cohort contained 2,654 nodules labeled benign, indeterminate, or malignant using median radiologist ratings. Patient-disjoint splits comprised 1,876 training, 382 validation, and 396 test nodules. Representations included one axial slice, five neighboring slices, and a local CT volume. Each track evaluated a randomly initialized baseline, momentum-contrast pretraining followed by fine-tuning, and supervised training with mild augmentation. Checkpoints were selected by validation Macro-F1. The 2D experiments used three seeds; 2.5D and 3D results were single-run estimates. A separate binary experiment excluded indeterminate nodules. Results: Augmented models achieved test Macro-F1 values of $0.6134\pm0.0075$ for 2D, $0.6537$ for 2.5D, and $0.5403$ for 3D; the 2D uncertainty denotes population standard deviation across seeds. Corresponding accuracies were $0.6128$ , $0.6869$ , and $0.5455$ . Augmentation improved Macro-F1 within every track, whereas the tested pretraining recipes reduced it. Binary Macro-F1 values were $0.8227$ , $0.8664$ , and $0.7554$ , respectively. Conclusion: Augmented 2.5D achieved the highest observed performance. Differences in preprocessing, optimization, capacity, and replication limit attribution to dimensionality. Labels reflect radiologist-assessed malignancy risk rather than pathology-confirmed cancer. Code and experimental results are available at https://github.com/CT-DeepLearning-Team/CT-2D-2.5D-3D-DeepLearning.

Figures & tables

Study	Representation and learning approach	Main difference from our primary endpoint
Cross-ViT (2025) [ 10 ]	Three-slice CNN–transformer feature fusion	Screened LUNA16 binary cohort; different reader aggregation; indeterminate excluded
MTST (2025) [ 11 ]	Multitask Swin Transformer; nodule-characteristic targets and GAN augmentation	Binary/multilevel targets exclude rating 3; additional supervision and generated images
LNMSNet (2026) [ 12 ]	Volumetric ResNet-18 with multiscale attention; segmentation and classification	Binary LIDC-IDRI target; voxel-mask supervision; separate external hospital evaluation
Present study	Residual networks with one slice, five slices, or a volume; supervised and MoCo recipes	Three median-rating groups; fixed patient-disjoint partition; unequal replication across tracks

Table 1: Qualitative context from recent pulmonary nodule studies. These methods were not reimplemented or evaluated on our test cohort; this table is a literature comparison, not an experimental leaderboard.

Figure 1: Study design. A common prepared cohort and patient assignments feed three representation tracks. Supervised labels are benign, indeterminate, and malignant. The SSL pool includes training-assigned patients only; the binary endpoint preserves patient assignments while removing indeterminate nodules.

Split	Patients	Benign	Indeterminate	Malignant	Total
Training	602	622	876	378	1,876
Validation	134	134	163	85	382
Test	134	117	187	92	396
Total	870	873	1,226	555	2,654

Table 2: Supervised cohort and fixed patient-disjoint assignments. Counts are nodules except in the patient column.

Setting	2D	2.5D	3D
Optimizer	AdamW	AdamW	AdamW
Batch size	64	8	4
Initial learning rate	$10^{-3}$	$10^{-3}$	$10^{-4}$
Weight decay	$0.05$	$10^{-4}$	$10^{-4}$
Maximum epochs	80	20	50
Early stopping patience	20	5	10

Table 3: Final supervised training settings. Track differences are retained rather than described as a common optimizer protocol.

Setting	2D	2.5D	3D
Queue size	4,096	512	4,096
Temperature	0.20	0.07	0.07
Batch size	128	8	4
Epochs	200	20	50
Optimizer	SGD	AdamW	AdamW
Initial learning rate	0.015	0.001	0.0003

Table 4: Pretraining configurations recorded in the repository. All use momentum

0.999

and projection output dimension 128.

Input	Model	Val. F1	Accuracy	Bal. acc.	Test Macro-F1	AUC
2D	A	.5982	.6002	.6061	$.6059\pm.0239$	.7565
2D	B	.6135	.5926	.5908	$.5957\pm.0086$	.7579
2D	C	.5962	.6128	.6110	$.6134\pm.0075$	.7595
2.5D	A	.5883	.6086	.6060	.6075	.7583
2.5D	B	.5516	.5126	.4914	.5125	.6922
2.5D	C	.5857	.6869	.6413	.6537	.7894

Table 5: Primary three-class results. 2D values are three-seed summaries; supplied SDs apply to Macro-F1. 2.5D and 3D are single runs. AUC is macro one-versus-rest.

Figure 2: Final test Macro-F1. (a) Three-class experiments A, B, and C; (b) separate binary experiment using the C training recipe. In 2D, bars show three-seed means, dots individual seeds, and error bars the population SD used in the stored summary. The 2.5D and 3D bars are single runs without uncertainty estimates. The dashed line in (a) is the always-indeterminate baseline. Binary and three-class scores evaluate different cohorts and targets.

Figure 3: Observed changes in test Macro-F1 compared to Model A: contrastive initialization (B–A) and mild augmentation (C–A). Bars show differences of reported three-seed means for 2D and single-run differences for 2.5D/3D; dots show same-seed downstream differences for 2D. No confidence interval or significance test for the changes is implied.

Input	Reference class	Precision	Recall	F1
2D	Benign	0.5467	0.5499	0.5421
2D	Indeterminate	0.6251	0.6346	0.6274
2D	Malignant	0.6950	0.6486	0.6708
2.5D	Benign	0.8261	0.3248	0.4663
2.5D	Indeterminate	0.6218	0.9144	0.7403
2.5D	Malignant	0.8400	0.6848	0.7545

Table 6: Model C class-specific test performance. 2D entries average the three seed-specific metrics; 2.5D and 3D entries are single-run estimates. Class names denote radiologist risk groups.

Figure 4: Row-normalized Model C confusion matrices. B, I, and M denote benign, indeterminate, and malignant risk groups. Labels within cells are counts and row percentages. The 2D counts pool three evaluations of the same 396 nodules (1,188 predictions), whereas the other matrices each contain one evaluation of 396 nodules. Color encodes the fraction within a reference class, enabling comparison despite unequal prediction counts.

Input	Val. F1	Accuracy	Bal. acc.	Test Macro-F1	AUC
2D	.8328	.8293	.8182	$.8227\pm.0024$	.8644
2.5D	.8583	.8708	.8614	.8664	.9215
3D	.7954	.7608	.7538	.7554	.8168

Table 7: Supplementary benign-versus-malignant results on 209 test nodules. Original patient assignments are retained.

Figure 5: Training and validation Macro-F1 for Models A and C. Planar 2D panels show all three seeds as separate curves; volumetric 3D panels show the single saved run. Solid curves denote training and dashed curves validation. Training metrics for C are measured with stochastic augmentation and therefore do not define a perfectly matched clean train–validation comparison. The 2.5D gap is described from its saved report rather than reconstructed as a numerical time series.

Medical Image Classification

Medical image classification uses computational models to identify diseases, grade severity, and support decisions from clinical images.

See all 505 papers

Evaluating Multi-Task Morphological Concept Learning for Pulmonary Nodule Malignancy Assessment in 3D CT
Sep 29, 2026Namitha NarayananRepresentation Learning Multi-Task Learning
Independent Researcher, Kerala, India
M3Net: A Macro-to-Meso-to-Micro Clinical-inspired Hierarchical 3D Network for Pulmonary Nodule Classification
May 12, 2026Jinyue Li, Yuzhou Yu, Jingjing Yang +73D Medical Imaging Hierarchical Representation Learning
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
Uncertainty-driven training for three-dimensional calibrated lung nodule classification
Sep 17, 2026Giuseppe Tripodi, Alessandro De Rosis, Saleh RezaeiraveshLung Nodule Detection Medical Image Classification
School of Engineering, The University of Manchester, Manchester (UK)

Sep 29, 2026J/K move · Enter open · S save

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

Namitha Narayanan

Independent Researcher, Kerala, India

Representation Learning Multi-Task Learning Medical Image Classification Computed Tomography

Open paper AskPDF arXiv

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.

Abstract

Medical Image Classification

Explore similar work

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

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

Uncertainty-driven training for three-dimensional calibrated lung nodule classification