PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluate the Malignancy of Pulmonary Nodules Using the 3D Deep Leaky Noisy-or Network

Fangzhou Liao, Ming Liang, Zhe Li, Xiaolin Hu, Sen Song

arXiv:1711.08324v1cs.CV

TL;DR

The paper addresses subject-level lung-cancer prediction when suspicious nodules do not individually determine malignancy and multiple nodules must be integrated. It uses a 3D detection-and-classification network with leaky noisy-or aggregation, achieving first place in the Data Science Bowl 2017 competition. The authors note that clinical use is limited because nodule growth over time is not considered.

  • Problem

    Lung-cancer diagnosis must detect suspicious nodules and integrate their morphology and malignancy evidence because nodule presence alone does not establish cancer.

  • Method

    A 3D region proposal network detects nodules, a classifier scores the top five, and a leaky noisy-or gate combines their cancer probabilities.

  • Results

    The proposed method won first place among 1972 teams in the Data Science Bowl 2017 competition, with test AUC 0.87 and test accuracy 81.42% at threshold 0.5.

  • Takeaways & Limitations

    Leaky noisy-or integration combines information from multiple lesions for disease prediction and can reduce the demand for highly accurate fine-scaled labels.

  • Takeaways & Limitations

    Clinical applicability is limited because the method does not consider nodule growth speed, which requires repeated scans and temporal alignment of nodules.

Abstract

from arXiv · show

Automatic diagnosing lung cancer from Computed Tomography (CT) scans involves two steps: detect all suspicious lesions (pulmonary nodules) and evaluate the whole-lung/pulmonary malignancy. Currently, there are many studies about the first step, but few about the second step. Since the existence of nodule does not definitely indicate cancer, and the morphology of nodule has a complicated relationship with cancer, the diagnosis of lung cancer demands careful investigations on every suspicious nodule and integration of information of all nodules. We propose a 3D deep neural network to solve this problem. The model consists of two modules. The first one is a 3D region proposal network for nodule detection, which outputs all suspicious nodules for a subject. The second one selects the top five nodules based on the detection confidence, evaluates their cancer probabilities and combines them with a leaky noisy-or gate to obtain the probability of lung cancer for the subject. The two modules share the same backbone network, a modified U-net. The over-fitting caused by the shortage of training data is alleviated by training the two modules alternately. The proposed model won the first place in the Data Science Bowl 2017 competition. The code has been made publicly available.

I. INTRODUCTION

Lung cancer diagnosis requires detecting suspicious nodules and integrating their morphology and cancer probabilities because nodule presence alone does not establish malignancy. The proposed system combines 3D nodule detection with leaky noisy-or aggregation for subject-level prediction.

  • Motivation: Pulmonary nodules vary in shape, can resemble normal tissues, and have a complicated relationship with cancer, making diagnosis difficult even for experienced doctors.Different doctors may give different malignancy predictions, and multiple nodules must be considered together.
  • Proposed approach: The system uses a 3D region proposal network to detect suspicious nodules and a classification stage to estimate subject-level cancer probability.The pipeline preprocesses the lungs, detects nodules, then scores and combines them.
  • Proposed approach: A leaky noisy-or gate combines cancer probabilities from multiple detected nodules while allowing cancer probability through a leakage term when no nodule explains the case.The model selects the top five suspicious nodules for classification and aggregation.
  • Training: The detection and classification networks share a modified backbone and are trained alternately to reduce over-fitting from limited training data.Extensive data augmentation is also used.
  • Contributions: The work proposes integrating a noisy-or gate into neural networks for multi-instance learning in computer-aided diagnosis.It also presents a volumetric one-stage end-to-end CNN for 3D object detection.
  • Result: The proposed method won first place among 1972 teams in the Data Science Bowl 2017 competition.The system uses classical preprocessing followed by neural-network detection and classification.

II. RELATED WORKS

Prior work addresses volumetric nodule detection and multi-instance learning through proposal generation, classifiers, and pooling layers. The paper positions its approach within limited 3D detection research and neural-network integration of noisy-or models.

  • Nodule detection: Extending 2D detection methods to 3D is constrained by GPU memory, motivating 2D proposal extraction followed by additional 3D proposal construction.The paper notes that direct 3D region proposal methods were limited in prior work.
  • Nodule detection: Volumetric nodule detection is commonly divided into proposal generation and false-positive reduction, with conventional and CNN-based methods addressing both subtasks.Earlier reviews evaluated conventional algorithms, while later work adopted larger datasets and CNNs.
  • Multiple instance learning: In multi-instance learning, a bag is positive if any instance is positive and negative if all instances are negative.Medical image analysis applications include pulmonary embolism and colon cancer detection.
  • Multiple instance learning: Deep multi-instance learning combines instance information through pooling layers such as max, mean, log-sum-exp, generalized-mean, concatenation, and noisy-or pooling.Pooling is the key component that aggregates instances within a neural-network framework.

D. Noisy-or model

Noisy-or models aggregate independent causes of an event, while the leaky variant permits residual probability when no observed factor explains it. The paper applies this framework to integrate nodule evidence across datasets with different annotation characteristics.

  • Noisy-or model: Noisy-or models infer disease probabilities by treating multiple factors as independent possible causes, and related work has used them in Bayesian medical diagnosis.Prior neural-network integrations were comparatively rare.
  • Datasets: The DSB dataset contains larger and anatomically different nodules than LUNA, including many nodules larger than 40 mm and on the main bronchus.Average nodule diameter is 13.68 mm in DSB versus 8.31 mm in LUNA.
  • Datasets: The authors remove LUNA annotations smaller than 6 mm and manually label DSB nodules to address dataset differences affecting detection.The DSB dataset provides subject-level cancer labels, whereas LUNA contains radiologist-annotated nodules.
  • Datasets: The cancer-classification stage is designed to remain robust to incorrect detections because manual nodule annotations may contain considerable noise.The authors state they lack professional lung-cancer diagnostic knowledge.

B. Preprocessing

Preprocessing isolates the lungs, expands masks to retain nodules attached to lung walls, suppresses surrounding bone and tissue, and converts CT data into network-ready UINT8 images.

  • Raw CT data are converted to Hounsfield Units, a standardized radiodensity scale with tissue-specific ranges.
  • Slices are Gaussian-filtered, thresholded at -600, and cleaned by removing small or highly eccentric connected components before lung extraction.Components smaller than 30 mm2 or with eccentricity greater than 0.99 are removed.
  • The lung mask is separated into left and right components before convex-hull computation, preventing unrelated tissues such as the heart and spine from entering the mask.
  • Convex hulls and 10-voxel dilation expand the two lung masks to include nodules attached to the outer lung wall, after which the masks are unioned.
  • For crescent-shaped lower-lung slices, the original mask is retained when its convex hull exceeds 1.5 times the mask area.
  • HU values are clipped to [-1200, 600], mapped to [0, 255], masked outside the lungs, and high-luminance bone is filled with tissue luminance 170.The resulting image is converted to UINT8; values above 210 in dilated surrounding space are also replaced with 170.

IV. 3D CNN FOR NODULE DETECTION

The nodule detector is a 3D RPN built on a modified U-Net, trained on patches because full-resolution lung scans exceed GPU memory and augmented to reduce over-fitting.

  • N-Net is a 3D CNN that uses a modified U-Net backbone and an RPN output layer to detect nodule and non-nodule proposals directly.Because the task has two classes, proposals are used as detection results without an additional classifier.
  • Full-image training is infeasible because a fine-resolution 3D lung scan can exceed the memory of mainstream GPUs.
  • Training therefore uses 128×128×128×1 3D patches, with 70% containing at least one nodule and 30% randomly cropped from lung scans.
  • The modified U-Net captures multi-scale information for variable nodule sizes, while its RPN output generates proposals directly.
  • The backbone combines a feedforward path of convolutions, residual blocks, and pooling with a feedback path of deconvolutions and combining units.
  • At each location, three anchors use lengths of 10, 30, and 60 mm, and the five regression outputs are (ô, d̂x, d̂y, d̂z, d̂r).A sigmoid is applied to the first regression value, while the remaining four use no activation.

C. Location information

The network incorporates normalized 3D proposal coordinates, while its loss combines classification and regression terms with regression restricted to positive anchors.

  • Each patch supplies normalized X, Y, and Z coordinate maps aligned with the output feature map as location information.
  • Anchor labels use IoU with target nodules, assigning positive and negative samples according to the stated overlap thresholds.
  • The classification loss uses the predicted sigmoid probability and binary label for each anchor.
  • Bounding-box regression predicts center offsets and side-length adjustment from the corresponding regression labels.
  • The total loss combines classification and regression losses, with regression applied only to positive samples because p = 1 there.Selected anchors are handled using positive-sample balancing and hard-negative mining.

E. Positive sample balancing

Training addresses imbalanced nodule sizes and abundant negatives by increasing sampling of large nodules and mining difficult negative examples.

  • Only one positive anchor is randomly selected for a large nodule to reduce correlation among its many corresponding positive anchors.
  • Sampling frequencies of large nodules are increased because small nodules substantially outnumber them and large nodules are stronger cancer indicators.
  • Hard-negative mining addresses the surplus of easily classified negative samples and the smaller subset whose appearances resemble nodules.
  • Negative samples are drawn into a random candidate pool, ranked by classification confidence, and the top n are retained as hard negatives.

G. Image splitting during testing

Testing splits lung scans into manageable parts, generates and suppresses nodule proposals, then evaluates the five highest-confidence nodules for case classification.

  • Lung scans are split into several 208 × 208 × 208 × 1 parts to overcome GPU memory limits during testing.
  • The detector outputs proposals containing nodule centers, radii, and confidence scores before non-maximum suppression removes overlaps.
  • Five proposals are selected by N-Net confidence during testing, while training samples proposals stochastically in proportion to confidence.
  • The classifier reuses N-Net features because limited training data makes an independent classification network prone to over-fitting.
  • Each selected proposal is converted into a 128-D feature from a cropped 96×96×96×1 patch, after which multiple-nodule integration methods produce a case score.
  • Feature combining uses a fully connected layer for each of five nodules, max-pools their 64-D features, and applies a sigmoid output for case cancer probability.

B. MaxP method

The MaxP method predicts cancer probability independently for each of the five selected nodules and uses the largest probability as the case prediction.

  • Each top-five nodule is processed by the same two-layer perceptron, whose sigmoid output is that nodule’s cancer probability.
  • The maximum nodule probability becomes the probability assigned to the entire case.
  • MaxP provides per-nodule interpretability but neglects interactions between nodules, such as combining two 50% probabilities into a higher case probability.
  • The leaky noisy-or model instead assumes nodules are independent causes of cancer and combines their individual probabilities, with a dummy nodule probability Pd.
  • Pd is learned automatically during training rather than manually tuned.
  • The leaky noisy-or model is the default case-classification model, called C-Net.

E. Training procedure

Training addresses limited data, memory constraints, and instability through augmentation, alternate detector-classifier updates, gradient clipping, and staged batch-normalization handling.

  • Classifier training uses pregenerated nodule bounding boxes, shared N-Net feature layers, and integration layers under a standard cross-entropy loss.
  • Data augmentation includes random flips, resizing between 0.75 and 1.25, arbitrary 3D rotations, and shifts in three directions.
  • Each training block alternates one detector training epoch with one classifier training epoch.
  • Gradient clipping normalizes gradients whose l2 norm exceeds one because small batches and outliers make training unstable.
  • Batch-normalization running statistics are unsuitable for both modules during alternate training because their inputs differ in patch size and task.
  • The three-stage procedure transfers detector weights, clips classifier gradients and freezes BN parameters, then alternates classification and detection with stored BN parameters.
  • On 198 validation cases containing 71 evaluated nodules, average recall was 0.8562 across 1/8 to 8 false positives per scan.
  • Evaluation showed that k = 5 was sufficient to capture most nodules.

B. Case classification

Case classification combines detected nodules into subject-level cancer predictions and evaluates training choices, integration methods, and benchmark performance.

  • Training schemes: Data augmentation prevented serious over-fitting, while alternate training significantly improved performance.The authors report that gradient clipping and batch-normalization freezing were not very useful in the compared schemes.
  • Training schemes: Clipping helped finetune stage A, alternate training further improved the model, and batch-normalization freezing improved the A→B→E scheme.The selected training procedure followed A→B→E.
  • Competition comparison: The proposed single model achieved the highest score among the top four competition teams, whose scores were very close.The method was evaluated using a rearranged training and validation split.
  • Nodule integration: Probability-based multi-nodule integration methods outperformed feature combining, with the Leaky Noisy-or model performing best.All compared integration models used alternate training configuration C.
  • Classification results: The evaluation reports predicted cancer-probability distributions and ROC curves for both training and test sets.The test-set AUC is 0.87, compared with 0.90 on training data.
  • Classification results: 85.96% and 81.42% classification accuracies were obtained on the training and test sets at a threshold of 0.5.At threshold 1, accuracies were 73.73% and 69.76%, respectively.

VII. DISCUSSION

The paper presents an end-to-end neural approach for lung cancer diagnosis and discusses its broader applicability, alongside limitations involving 3D computation, training data, and clinical temporal information.

  • VII. DISCUSSION: The leaky noisy-or network may apply to medical image-analysis tasks where multiple lesions provide uncertain evidence about disease.It can integrate information from different lesions and reduce the demand for highly accurate fine-scaled labels.
  • VII. DISCUSSION: 3D CNN deployment is constrained by memory, limiting running speed, batch size, and model depth.The authors address these constraints with a shallower network and image patches instead of whole-image inputs.
  • VII. DISCUSSION: The model remains vulnerable to overfitting because 3D CNNs have substantially more parameters than comparable 2D CNNs.Data augmentation and alternate training are used to mitigate this problem.
  • VII. DISCUSSION: The 1700-case training set may not cover the full variation of pulmonary nodules, motivating more samples and incorporation of nodule-segmentation labels.The paper identifies both strategies as potential ways to improve performance.
  • VII. DISCUSSION: The competition task omits nodule growth speed, an important clinical factor that requires repeated scans, temporal alignment, and detection of small nodules.The authors note that the proposed method does not pursue high detection accuracy for small nodules but could be modified for this purpose.
Loading 1711.08324v1…