PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Bahodir Nazarov, Zineb Gharib, Fatima Zazoul

arXiv:2610.35960v1cs.LGcs.CV

TL;DR

Whether added spatial context improves pulmonary-nodule malignancy-risk classification remains an empirical question. The paper compares 2D, 2.5D, and 3D residual-network tracks and finds the highest observed three-class test Macro-F1 with augmented 2.5D, while implementation differences limit attribution to dimensionality.

  • Problem

    Single-slice, neighboring-slice, and volumetric CT representations provide different spatial information, and whether richer representations improve held-out performance remains unresolved.

    Source: §1 · p. 1

  • Method

    The study compares 2D, 2.5D, and 3D residual-network tracks with direct supervision, momentum-contrast initialization, mild augmentation, and a separate binary experiment.

    Source: §1 · p. 2, §3.6 · p. 6

  • Results

    Augmented 2.5D achieved the highest observed three-class test Macro-F1 (0.6537) on the fixed LIDC-IDRI-derived patient-disjoint cohort.

    Source: §6 · p. 13

  • Takeaways & Limitations

    Full-volume modeling did not outperform planar alternatives under the evaluated data and training constraints, without establishing universal 2.5D superiority or clinical diagnostic performance.

    Source: §6 · p. 13

    One fixed patient partition, no external dataset, unequal repeated-seed reporting, correlated nodules, and preprocessing and optimization differences limit claims of statistical superiority or dimensionality-only effects.

    Source: §5.4 · p. 12

Abstract

from arXiv · show

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.

1 Introduction

A single slice, neighboring slices, and a local volume provide different spatial information, but whether richer representations improve held-out performance is an empirical question. The study compares these representations and training strategies while noting that its implemented tracks are not a strict single-variable ablation.

1 Introduction

  • Whether a single axial image, neighboring slices, or a local 3D volume yields better held-out performance remains an empirical question.A single image may omit through-plane variation, neighboring slices provide limited depth context, and volumetric convolutions learn across three spatial axes.Source: §1 · p. 1
  • LIDC-IDRI reader ratings support image-based risk characterization, but the resulting classes represent reader assessments rather than histopathology-confirmed disease.The paper therefore uses the term malignancy-risk classification.Source: §1 · p. 1
  • The study compares three residual-network representations, contrastive initialization, and mild augmentation on a common cohort, while retaining uncertain ratings and a separate binary experiment.The contributions also include error analysis and documented implementation differences.Source: §1 · p. 2
  • The implemented model tracks are not a strict single-variable ablation because preprocessing, optimization, capacity, and replication differ across representations.The study reports these differences explicitly.Source: §1 · p. 2

2 Related Work

Prior pulmonary-nodule work incorporates morphology, spatial context, attention, and auxiliary supervision, but differs in representation, task, or label policy. The present study contributes comparative evidence on three residual-network representations, not a new architecture.

2 Related Work

  • Shen et al. used multi-crop features for malignancy suspiciousness, whereas this study's 2.5D input is a fixed neighboring-slice stack.The approaches address contextual information but are not interchangeable representations.Source: §2 · p. 2
  • Setio et al.'s multi-view work targets false-positive reduction in nodule detection, so it is not a directly comparable baseline for malignancy-risk Macro-F1.Published scores under different labels, cohort selection, and partitions are also not directly comparable.Source: §2 · p. 2
  • Related studies use neighboring slices, attention, or auxiliary tasks, but their binary or multilevel experiments often exclude indeterminate ratings, unlike this study's primary cohort.The cited work motivates follow-up directions without resolving the best representation under this study's labels and splits.Source: §2 · p. 2
  • The study contributes comparative evidence on three residual-network representations with attention to uncertain ratings, pretraining transfer, and supervised regularization, rather than a new architecture.Source: §2 · p. 2

3 Materials and Methods

The retrospective study compares implemented model tracks on a fixed LIDC-IDRI-derived cohort with patient-disjoint assignments. Median reader ratings define the primary three-class labels, while a separate binary experiment removes indeterminate nodules.

3.1 Study design and source provenance

  • The retrospective computational study compares three model tracks on a fixed LIDC-IDRI-derived cohort, with internal patient-disjoint evaluation and no external validation cohort.The accompanying repository versions methods and results, and the design documents differences that limit cross-representation causal inference.Source: §3.1 · p. 3

3.2 Dataset preparation and cohort

  • The preparation pipeline produced 2,654 eligible supervised nodules from 1,010 patients, and fixed assignments keep each patient's eligible nodules together.The preparation counts refer to distinct pipeline stages, not independent scans or supervised examples; Table 2 summarizes the fixed patient-disjoint assignments.Source: §3.2 · p. 3, Table 2 · p. 4

3.3 Reference labels

  • Median available reader ratings define benign, indeterminate, and malignant classes, while the separate binary experiment removes indeterminate nodules and preserves patient assignments.The primary endpoint retains indeterminate ratings as a substantial part of the cohort and an area of reader uncertainty.Source: §3.3 · p. 5

3.4 Image representations and actual preprocessing

The three representations use different CT inputs and preprocessing, and the volumetric model also differs in capacity and computational demands. These implementation differences qualify cross-track comparisons.

3.4 Image representations and actual preprocessing

  • The 2D loader uses axial slice 52, 2.5D selects five slices around the mask-based center, and 3D supplies the complete scaled crop.The source crops have shape (104, 72, 80) at 1-mm isotropic spacing, and the common window maps retained intensities to [0, 1].Source: §3.4 · p. 5
  • Shared data and an HU window do not make preprocessing identical: 2.5D uses mask-dependent centering, 2D fixed centering, and 2D adds standardization.These are characteristics of the evaluated implementations and matter when interpreting cross-track differences.Source: §3.4 · p. 5

3.5 Network architectures

  • The final volumetric model is a lightweight 3D ResNet-18 with 2,073,747 three-class parameters, and its larger input and spatial operations create a different computational and optimization problem.Equal architectural family does not imply equal capacity, receptive fields, or training cost.Source: §3.5 · p. 5

3.6 Experimental conditions and optimization

The study compares three supervised model conditions within each representation, selecting checkpoints by validation Macro-F1. It also tests momentum-contrast initialization, with track-specific pretraining configurations.

3.6 Experimental conditions and optimization

  • Model A trains from random initialization without augmentation, Model B fine-tunes a momentum-contrast-initialized encoder, and Model C trains from random initialization with mild augmentation.The binary experiment uses the Model C recipe with a two-output head.Source: §3.6 · p. 6
  • Every track selects the checkpoint with maximum validation Macro-F1 and applies early stopping after the stated number of epochs without improvement.Test metrics are computed at the selected checkpoint.Source: §3.6 · p. 6

3.7 Momentum-contrast pretraining

  • MoCo pretraining compares two augmented views of the same nodule against a queue of representations from other samples [3].The key encoder follows the query encoder by an exponential moving average; the projection head is discarded when transferring the encoder to classification.Source: §3.7 · p. 6
  • Pretraining configurations use momentum 0.999 and projection output dimension 128, but the substantially different pretraining budgets test track-specific recipes rather than a uniform SSL configuration.Risk labels are ignored during pretraining, which uses only training-assigned patients.Source: §3.7 · p. 6, Table 4 · p. 7

3.8 Training augmentation

Training augmentation uses representation-specific transformations, while evaluation reports test Macro-F1 and additional classification metrics. Uncertainty reporting differs across tracks.

3.8 Training augmentation

  • The 2D mild recipe includes flips, affine rotation up to 15°, scale 0.9–1.1, intensity shift ±0.05, and Gaussian noise with standard deviation 0.02.Augmentation occurs before standardization.Source: §3.8 · p. 7
  • The 2.5D recipe applies flips and intensity/noise settings jointly across five channels, while the 3D recipe uses rotations up to 5° per axis, translations up to three voxels, and milder intensity/noise changes.The 2.5D recipe omits the 2D affine transform; values in the 3D recipe are clipped back to [0, 1].Source: §3.8 · p. 7

3.9 Evaluation and uncertainty

  • Evaluation also reports accuracy, per-class precision and recall, confusion matrices, and macro one-versus-rest ROC-AUC for three-class tasks.The majority-class test predictor has accuracy 187/396 = 0.4722, balanced accuracy 1/3, and Macro-F1 approximately 0.2138.Source: §3.9 · p. 7
  • The 2D values aggregate three downstream seeds, whereas 2.5D and 3D lack corresponding repeated-seed estimates.The 2D standard deviations describe seed variation, not confidence intervals on representation differences.Source: §3.9 · p. 7

4 Results

Among the tested three-class Model C runs, 2.5D had the highest observed overall scores. Augmentation improved Macro-F1 within each track, while class-specific tradeoffs and differences in replication qualify comparisons.

4.1 Three-class performance

  • 2.5D Model C had the highest observed Macro-F1 (0.6537), accuracy (0.6869), balanced accuracy (0.6413), and ROC-AUC (0.7894).Model A had similar Macro-F1 for 2D (0.6059) and 2.5D (0.6075), while 3D was lower (0.4928).Source: Table 5 · p. 8, §4.1 · p. 8

4.2 Augmentation and contrastive initialization

  • Model C’s test Macro-F1 changes versus Model A were +0.0075 for 2D, +0.0462 for 2.5D, and +0.0475 for 3D.These positive changes support a regularization interpretation for the 2.5D run, but its augmented training measurements are not identical to clean evaluation measurements.Source: §4.2 · pp. 8–9, Figure 3 · p. 9
  • MoCo decreased downstream Macro-F1 relative to Model A by −0.0102 for 2D, −0.0950 for 2.5D, and −0.0198 for 3D.For 3D, AUC rose from 0.6564 to 0.6692 under Model B even as Macro-F1 decreased.Source: §4.2 · p. 9

4.3 Class-specific errors

  • 2.5D Model C correctly classified 38/117 benign, 171/187 indeterminate, and 63/92 malignant nodules, despite benign recall of only 0.3248.It predicted 275/396 nodules as indeterminate, versus 187 true indeterminate labels.Source: §4.3 · pp. 9–10, Figure 4 · p. 10
  • 3D Model C had higher benign recall but lower indeterminate recall than 2.5D, so its lower overall Macro-F1 did not mean worse recall for every class.The 3D recalls were 0.3761, 0.5882, and 0.6739 for benign, indeterminate, and malignant, respectively.Source: §4.3 · p. 10

4.4 Binary sensitivity analysis

  • 2.5D had the highest observed binary Macro-F1 (0.8664), but this endpoint removed indeterminate nodules and evaluated a different cohort and target.The separate binary test set contained 209 nodules.Source: Table 7 · p. 10, §4.4 · p. 10

4.5 Replication and learning behavior

  • The 2D Model C seed values were 0.6033, 0.6211, and 0.6158; the average A–C improvement did not occur for every seed.Their nodule-bootstrap intervals describe individual seed predictions, not an interval around the three-seed mean; 2.5D and 3D were single runs.Source: §4.5 · p. 11, Figure 5 · p. 11

5 Discussion

Augmented 2.5D achieved the strongest observed three-class result, but the comparison does not establish that five slices are intrinsically optimal or that 3D information is harmful. The discussion also limits conclusions about augmentation, pretraining, clinical meaning, and comparisons with recent methods.

5.1 Spatial context and generalization

  • Augmented 2.5D achieved the strongest observed three-class result, but the ranking does not establish that five slices are intrinsically optimal or that 3D information is harmful.The authors describe this as a hypothesis about the implemented tracks; another 3D architecture, training budget, crop extent, or regularization recipe might change the ranking.Source: §5.1 · pp. 11–12

5.2 Regularization and self-supervised transfer

  • Mild augmentation had a consistent positive A–C direction, while negative A–B changes apply to the tested pretraining recipes, not self-supervised learning generally.Transformations differed by representation, and only 2D had repeated-seed results, so the experiments do not quantify a universal augmentation effect.Source: §5.2 · p. 12

5.3 Uncertainty and interpretation of the endpoint

  • The binary analysis suggests intermediate risk contributes to a difficult three-class decision, but cohort size, class balance, and output classes also differ.The endpoint measures agreement with aggregated reader ratings, not cancer sensitivity, specificity, prospective benefit, or clinical utility.Source: §5.3 · p. 12

5.4 Limitations and next experiments

  • A fixed patient partition, no external dataset, unequal seed replication, within-patient nodule correlation, and differing implementation choices constrain attribution and claims of superiority.The authors recommend harmonized processing and training, matched repeated runs, and patient-level paired differences on an untouched evaluation set; final design decisions may also have prior test-observation exposure.Source: §5.4 · pp. 12–13

5.5 Scope of comparison with recent methods

  • Without matched evaluations of Cross-ViT, MTST, or LNMSNet, these residual-network results neither establish superiority over recent methods nor isolate input dimensionality.Their cohorts, risk aggregation, excluded ratings, supervision, and test partitions differ from this study’s.Source: §5.5 · p. 13

6 Conclusion

On a fixed LIDC-IDRI-derived patient-disjoint cohort, augmented 2.5D had the highest observed three-class test Macro-F1. The findings are limited to the evaluated data and training constraints and do not establish universal superiority or clinical diagnostic performance.

6 Conclusion

  • 0.6537 was the highest observed three-class test Macro-F1, achieved by augmented 2.5D on the fixed LIDC-IDRI-derived patient-disjoint cohort.The findings do not establish universal 2.5D superiority or clinical diagnostic performance.Source: §6 · p. 13
  • Mild augmentation improved reported mean or single-run Macro-F1 across all tracks, while the tested MoCo recipes did not.Source: §6 · p. 13
  • Removing indeterminate nodules produced substantially higher scores on a separate binary endpoint.Source: §6 · p. 13

Data and code availability

The repository provides code and experimental materials, while the imaging data and reader annotations are distributed separately.

Data and code availability

Statements and Declarations

The declarations report no funding or conflict of interest; the study is a secondary analysis of publicly released LIDC-IDRI data.

Funding

Competing interests

Ethical considerations

  • The study analyzes publicly released LIDC-IDRI images and annotations without recruiting participants or performing additional imaging or intervention.Source: Ethical considerations · p. 14
Loading 2610.35960v1…