[Paper Review] 3D G-CNNs for Pulmonary Nodule Detection

Marysia Winkels, Taco Cohen|arXiv (Cornell University)|Apr 12, 2018

Lung Cancer Diagnosis and Treatment24 references75 citations

TL;DR

The paper demonstrates that 3D group-equivariant CNNs (G-CNNs) substantially improve false positive reduction for pulmonary nodule detection and are roughly 10x more data-efficient than standard 3D CNNs.

ABSTRACT

Convolutional Neural Networks (CNNs) require a large amount of annotated data to learn from, which is often difficult to obtain in the medical domain. In this paper we show that the sample complexity of CNNs can be significantly improved by using 3D roto-translation group convolutions (G-Convs) instead of the more conventional translational convolutions. These 3D G-CNNs were applied to the problem of false positive reduction for pulmonary nodule detection, and proved to be substantially more effective in terms of performance, sensitivity to malignant nodules, and speed of convergence compared to a strong and comparable baseline architecture with regular convolutions, data augmentation and a similar number of parameters. For every dataset size tested, the G-CNN achieved a FROC score close to the CNN trained on ten times more data.

Motivation & Objective

• Motivate data efficiency in medical image analysis for pulmonary nodule detection due to limited annotated data.
• Propose and evaluate 3D group-equivariant convolutions to improve learning with rotations and reflections.
• Compare G-CNNs to a translational baseline across varying training set sizes.
• Assess convergence speed and data efficiency of 3D G-CNNs on NLST/LIDC/IDRI datasets.

Proposed method

• Replace standard Conv3D with GConv3D for selected symmetry groups (D4, D4h, O, Oh).
• Generate transformed filter banks by applying group transformations; shuffle orientation channels via precomputed permutation rho(h).
• Balance augmented networks to keep parameter counts comparable to the baseline by scaling outputs with sqrt(|H|).
• Train using cross-entropy with Adam, Xavier initialization, and a single data augmentation scheme (rotations, reflections, translations, scaling, noise).
• Evaluate on NLST/LIDC/IDRI with 30–30,000 samples, using FROC analysis and a fixed evaluation protocol (ANODE09/LUNA16-style).

Experimental results

Research questions

• RQ1 How do 3D G-CNNs with roto-reflection groups perform in false positive reduction for pulmonary nodules compared to standard 3D CNNs?
• RQ2 Are 3D G-CNNs more data-efficient and faster to converge than translational CNNs across varying training set sizes?
• RQ3 What is the impact of different 3D symmetry groups (D4, D4h, O, Oh) on performance, especially in small data regimes?

Key findings

• G-CNNs outperform the translational baseline across all dataset sizes.
• G-CNNs trained on N data can match or exceed the baseline trained on 10×N data, indicating ~10× data efficiency.
• FROC scores improve with all G-convolution groups, with notable gains in small data regimes.
• G-CNNs converge faster than baseline, often requiring far fewer epochs to reach comparable validation losses.
• Among groups, octahedral groups (O, Oh) and rectangular cuboid groups (D4, D4h) show different performance patterns, particularly at very small data sizes.
• Malignancy sensitivity in top true positives tends to be higher for G-CNNs than for the baseline.