Source-linked AI summary
A New 2.5D Representation for Lymph Node Detection using Random Sets of Deep Convolutional Neural Network Observations
Holger R. Roth, Le Lu, Ari Seff, Kevin M. Cherry, Joanne Hoffman, Shijun Wang, Jiamin Liu, Evrim Turkbey, Ronald M. Summers
TL;DR
The paper addresses false positives in challenging CT lymph-node detection, where direct 3D CNNs face computational and training-data constraints. It uses random 2.5D CNN observations and reports improved sensitivity at clinically relevant false-positive rates in mediastinal and abdominal datasets.
Problem
Automated lymph-node detection in CT is difficult because of low contrast and variation in lymph-node size, pose, shape, and location.
Method
The method resamples each 3D volume of interest into randomly translated, rotated, and scaled 2D orthogonal views, classifies them with a CNN, and averages their probabilities.
Results
70% sensitivity at 3 FP/vol. in the mediastinum and 83% at 3 FP/vol. in the abdomen, compared with 52.9% at 3.1 FP/vol. and 70.5% at 13.0 FP/vol., respectively, in cited prior work.
Takeaways & Limitations
The results support random sets of CNN observations as an effective 2.5D strategy for reducing false positives in 3D medical-image computer-aided detection.
Takeaways & Limitations
More sophisticated methods for fusing CNN view labels remain future work.
Abstract
from arXiv · showhide
Automated Lymph Node (LN) detection is an important clinical diagnostic task but very challenging due to the low contrast of surrounding structures in Computed Tomography (CT) and to their varying sizes, poses, shapes and sparsely distributed locations. State-of-the-art studies show the performance range of 52.9% sensitivity at 3.1 false-positives per volume (FP/vol.), or 60.9% at 6.1 FP/vol. for mediastinal LN, by one-shot boosting on 3D HAAR features. In this paper, we first operate a preliminary candidate generation stage, towards 100% sensitivity at the cost of high FP levels (40 per patient), to harvest volumes of interest (VOI). Our 2.5D approach consequently decomposes any 3D VOI by resampling 2D reformatted orthogonal views N times, via scale, random translations, and rotations with respect to the VOI centroid coordinates. These random views are then used to train a deep Convolutional Neural Network (CNN) classifier. In testing, the CNN is employed to assign LN probabilities for all N random views that can be simply averaged (as a set) to compute the final classification probability per VOI. We validate the approach on two datasets: 90 CT volumes with 388 mediastinal LNs and 86 patients with 595 abdominal LNs. We achieve sensitivities of 70%/83% at 3 FP/vol. and 84%/90% at 6 FP/vol. in mediastinum and abdomen respectively, which drastically improves over the previous state-of-the-art work.
1 Introduction
LN detection supports disease staging and treatment but is difficult because nodes vary in appearance, location, size, pose, and shape, while 3D modeling is constrained by limited annotated data and high dimensionality.
- Motivation: Accurate LN detection and segmentation support disease staging, prognosis, therapy choice, and follow-up for conditions including lung cancer, lymphoma, and inflammation.LNs are considered enlarged when their smallest axial diameter exceeds 10 mm.
- Prior work: 3D CADe methods select and combine features from approximately 50 thousand 3D Haar-like features using boosting-based classifiers.These methods directly model volumetric CT information.
- Challenge: Modeling complex 3D LN structures is non-trivial because annotated training data are limited and the feature space is intrinsically high-dimensional.LNs also exhibit substantial within-class variation in appearance and location.
- Approach: The proposed work investigates CNNs for false-positive reduction and introduces a 2.5D representation intended to facilitate 3D object detection by classification.The approach builds on a high-sensitivity LN CADe stage.
2 Methods
The method converts candidate-centered 3D VOIs into randomized three-channel 2D views and classifies them with a CNN, averaging view probabilities to obtain a VOI score.
- Candidate generation: The preliminary CADe systems generate LN candidates from mediastinal and abdominal CT volumes using region-specific anatomical, voxel-level, SVM, and random-forest processing.The mediastinal system uses lung segmentation and anatomical spatial priors, whereas the abdominal system produces voxel-level LN predictions with a random forest.
- 2.5D representation: Each VOI is represented as an RGB image whose red, green, and blue channels contain axial, coronal, and sagittal slices through a candidate-centered location.This jointly uses three orthogonal views rather than separately training one CNN per slice orientation.
- 2.5D representation: VOIs are resampled at multiple physical scales while retaining a fixed voxel count, with random translations and rotations augmenting the training views.The perturbations vary the sampled representation around each candidate.
- CNN aggregation: The CNN classifies the randomized three-channel views, and averaging their probabilities produces a final classification probability for each VOI.This random-set aggregation decomposes volumetric information into orthogonal 2D views at multiple reformatted orientations.
- CNN aggregation: The 2.5D resampling avoids direct 3D CNN training, reducing computational burden and helping address high dimensionality and limited medical-imaging training samples.Random resampling also increases training-data variation and helps reduce overfitting.
- CNN architecture: The CNN uses convolutional, max-pooling, locally fully-connected, DropConnect, and final two-way softmax layers.DropConnect is used as a training regularizer, while GPU acceleration and rectified linear units support efficient training.
3 Evaluation and Results
The evaluation uses patient-level cross-validation on mediastinal and abdominal LN candidates, showing improved classification sensitivity and abdominal AUC after CNN-based false-positive reduction.
- Datasets: 388 mediastinal LNs from 90 patients and 595 abdominal LNs from 86 patients were labeled as positive examples.The datasets provide the positive LN cases for evaluating the CNN-based 2.5D module.
- Evaluation setup: 100% sensitivity was assumed at candidate generation by injecting labeled LNs, while 3208 mediastinal and 3484 abdominal false-positive detections supplied negative examples.Patients were randomly divided at the patient level for 3-fold cross-validation.
- Results: At 3 FP/vol. with N = 100, sensitivity increased from 55% to 70% in the mediastinum and from 30% to 83% in the abdomen over existing CADe systems.The comparison concerns the CNN-based false-positive reduction stage.
- Results: AUC in the abdomen increased from 0.76 to 0.942 using the proposed false-positive reduction approach.AUC comparison was unavailable for the mediastinal dataset.
- Joint training: Joint training on mediastinal and abdominal candidates improved mediastinal sensitivity by approximately 10% to approximately 80% at 3 FP/vol., while abdominal sensitivity remained stable.The result indicates a benefit from combining the two candidate domains for mediastinal classification.
4 Discussion and Conclusions
The proposed 2.5D random-set CNN observers reduce false positives in lymph-node CADe and improve sensitivity at clinically relevant FP/vol. rates. Performance also benefits from joint training on mediastinal and abdominal nodes, while simple averaging remains effective.
- Method and contribution: The 2.5D approach uses random scales, translations, and rotations around CADe detections to reduce false positives and alleviate overfitting.It decomposes 3D volumes into random reformatted 2D views processed by CNN observers.
- Performance: 70% sensitivity at 3 FP/vol. in the mediastinum improves on 52.9% at 3.1 FP/vol. reported by Feulner et al. (2013).The authors note that direct comparisons are difficult because no common datasets are available.
- Performance: 83% sensitivity at 3 FP/vol. in the abdomen exceeds 70.5% at 13.0 FP/vol. reported by Nakamura et al. (2013).The abdominal result assumes approximately 100% sensitivity at the LN candidate-generation stage.
- Evaluation: AUC and FROC show significant sensitivity improvement at clinically relevant FP/vol. rates, with curves evaluated for varying numbers of random CNN observers.Figure 4 reports 3-fold cross-validation in 90 mediastinal and 86 abdominal patients.
- Training data: Joint training on mediastinal and abdominal lymph nodes improves performance, supporting the benefit of larger, more varied, comprehensive datasets.This finding is presented as coherent with prior computer-vision literature.
- Evaluation: Sparsely weighted fusion does not improve over the simple average of the CNN predictions.The authors describe the simple average as effective and efficient and propose more sophisticated label fusion for future work.