PaperlayerView exact revision on arXiv

Source-linked AI summary

Computer-aided diagnosis of lung nodule using gradient tree boosting and Bayesian optimization

Mizuho Nishio, Mitsuo Nishizawa, Osamu Sugiyama, Ryosuke Kojima, Masahiro Yakami, Tomohiro Kuroda, Kaori Togashi

arXiv:1708.05897v2cs.CV

TL;DR

False positives in low-dose CT screening can lead to unnecessary follow-up, motivating interest in CADx for benign–malignant nodule differentiation. The study evaluated a lung-nodule CADx system using LBP-derived features, SVM or XGBoost, and parameter optimization. XGBoost outperformed SVM, while TPE was more efficient than random search; the best averaged XGBoost AUC was 0.896.

  • Problem

    False positives in low-dose CT screening can lead to unnecessary follow-up, while CADx is used to differentiate benign from malignant nodules.

    Source: Introduction · p. 3

  • Method

    The system used local-binary-pattern feature vectors with SVM or XGBoost and compared parameter-optimization approaches.

    Source: Introduction · p. 4, Machine learning algorithm · p. 6, Discussion · p. 9

  • Results

    XGBoost was better than SVM for classifying lung nodules, with a best averaged AUC of 0.896 under leave-one-out cross-validation.

    Source: Discussion · pp. 10–11

  • Takeaways & Limitations

    Bayesian optimization was more efficient than random search for optimizing parameters of SVM and XGBoost.

    Source: Discussion · p. 12

    The number of lung nodules was relatively small, limiting the study's sample scope.

    Source: Discussion · p. 11

Abstract

from arXiv · show

We aimed to evaluate computer-aided diagnosis (CADx) system for lung nodule classification focusing on (i) usefulness of gradient tree boosting (XGBoost) and (ii) effectiveness of parameter optimization using Bayesian optimization (Tree Parzen Estimator, TPE) and random search. 99 lung nodules (62 lung cancers and 37 benign lung nodules) were included from public databases of CT images. A variant of local binary pattern was used for calculating feature vectors. Support vector machine (SVM) or XGBoost was trained using the feature vectors and their labels. TPE or random search was used for parameter optimization of SVM and XGBoost. Leave-one-out cross-validation was used for optimizing and evaluating the performance of our CADx system. Performance was evaluated using area under the curve (AUC) of receiver operating characteristic analysis. AUC was calculated 10 times, and its average was obtained. The best averaged AUC of SVM and XGBoost were 0.850 and 0.896, respectively; both were obtained using TPE. XGBoost was generally superior to SVM. Optimal parameters for achieving high AUC were obtained with fewer numbers of trials when using TPE, compared with random search. In conclusion, XGBoost was better than SVM for classifying lung nodules. TPE was more efficient than random search for parameter optimization.

Introduction

The introduction motivates CADx for benign–malignant nodule differentiation amid false positives in low-dose CT screening. It frames the study around XGBoost and parameter-optimization strategies for lung-nodule classification.

Introduction

  • 96.4% of positive low-dose CT screening results in NLST were false positives, which can prompt unnecessary follow-up imaging or invasive procedures.The passage lists follow-up CT, positron emission tomography, and invasive procedures as possible consequences.Source: Introduction · p. 3
  • CADx differentiates benign from malignant nodules, and the authors expect it could help reduce false positives in low-dose CT screening.Source: Introduction · p. 3
  • XGBoost was applied to lung-nodule CADx to address a gap in prior investigations of gradient tree boosting for this task.The authors state that, to their knowledge, no earlier study had investigated its usefulness in lung-nodule CADx.Source: Introduction · p. 3
  • Grid search becomes computationally infeasible as the number of parameters increases, motivating consideration of random search and Bayesian optimization.Source: Introduction · pp. 3–4
  • The study evaluated XGBoost against SVM and compared random search with Bayesian optimization for parameter optimization.The stated aims were to assess XGBoost's usefulness and the effectiveness of the two optimization approaches.Source: Introduction · p. 4
  • The system calculated feature vectors from a hand-crafted imaging feature, a variant of local binary pattern.Source: Introduction · p. 4

Methods

The study used anonymized public CT data and combined lung-nodule sets from LUNGx Challenge and NSCLC Radiogenomics for CADx development and evaluation.

Methods

  • The study used anonymized data from public databases, for which Japanese regulations did not require institutional review board approval.Source: Methods · p. 4

CT images

  • 99 nodules—62 cancers and 37 benign nodules—were combined from the LUNGx Challenge and NSCLC Radiogenomics datasets.The images were chest CT images obtained through The Cancer Imaging Archive.Source: CT images · pp. 4–5

Image preprocessing

The pipeline resampled CT volumes, cropped a fixed bounding box around each nodule, and converted its texture into a one-dimensional LBP-TOP feature vector.

Image preprocessing

  • CT volumes were resampled to 1 × 1 × 1 mm, and a 64 × 64 × 64 bounding box was cropped around each nodule.For NSCLC Radiogenomics, two board-certified radiologists visually validated the nodule center; image areas outside the box were not assessed.Source: Image preprocessing · p. 5

Calculation of a feature vector

  • Uniform-pattern and rotation-invariant LBP were used because naïve LBP cannot handle large P values; the authors say both variants can enhance feature-vector robustness.Source: Calculation of a feature vector · p. 6
  • LBP-TOP calculated 2D LBP on the XY, XZ, and YZ planes, converted the results to histograms, and concatenated them into a 1D feature vector.Texture information on other 3D planes was ignored.Source: Calculation of a feature vector · pp. 5–6

Machine learning algorithm

The CADx system trained SVM or XGBoost on LBP-TOP feature vectors and labels. Parameter optimization covered controls for both classifiers and LBP-TOP.

Machine learning algorithm

  • SVM with a radial-basis-function kernel or XGBoost was trained on LBP-TOP feature vectors and their corresponding labels.XGBoost uses gradient tree boosting, which is invariant to feature-vector scaling and can determine higher-order interactions.Source: Machine learning algorithm · p. 6

Parameters

  • The parameter space included SVM C and γ, XGBoost controls, and the LBP-TOP parameters R and P.The listed XGBoost controls include eta, max_depth, min_child_weight, and gamma.Source: Parameters · p. 7

Parameter optimization

The optimization objective was to minimize validation loss across K-fold cross-validation for SVM or XGBoost. TPE and random search were compared for finding parameters.

Parameter optimization

  • The validation loss measured performance on validation data after training the selected algorithm with parameters θ.The passages denote the algorithm as A and the training and validation data as Dtrain and Dvalid.Source: Parameter optimization · p. 8
  • Parameter optimization minimized the black-box validation-loss function across the K folds.The algorithm was trained on each fold’s training data and validated on its validation data.Source: Parameter optimization · p. 8
  • TPE was used for Bayesian optimization and compared with random search at 10, 100, 200, and 1000 trials.The compared algorithms were SVM and XGBoost.Source: Parameter optimization · p. 8

Statistical analysis

Leave-one-out cross-validation was used to optimize and evaluate the CADx system. Its probabilistic outputs were assessed with AUC and accuracy, averaged across 10 calculations for each trial count.

Statistical analysis

  • Leave-one-out cross-validation was used for both parameter optimization and CADx performance evaluation.Validation loss under leave-one-out cross-validation was used to optimize parameters.Source: Statistical analysis · p. 8
  • The optimized CADx system’s probabilistic lung-cancer outputs were evaluated using ROC AUC and accuracy.The system output probabilities of lung cancer to calculate AUC.Source: Statistical analysis · p. 8
  • AUC and accuracy were each calculated 10 times for every trial count, and their averages were reported.The system’s evaluation outline is shown in Fig. 1.Source: Statistical analysis · p. 8

Results

XGBoost generally outperformed SVM, and each algorithm’s best averaged AUC was obtained with TPE at 1000 trials. TPE’s advantage over random search varied by algorithm and trial count.

Results

  • XGBoost was generally superior to SVM in the reported averages: its best averaged AUC was 0.896 at 1,000 trials, while SVM’s best averaged AUC was 0.850 using TPE.Source: Results · p. 9
  • XGBoost’s averaged AUC was higher with TPE than random search at 100, 200, and 1000 trials.The supplied results do not specify the size of these differences.Source: Results · p. 9
  • For SVM, the difference in averaged AUC between random search and TPE was minimal at 10, 200, and 1000 trials; at 10 trials, this was also true for XGBoost.Source: Results · p. 9

Discussion

The discussion concludes that XGBoost outperformed SVM and Bayesian optimization was more efficient than random search, while emphasizing preliminary results and limits on validation and scope.

Discussion

  • 0.896 was the best averaged AUC under leave-one-out cross-validation using XGBoost and TPE; XGBoost's AUC was better than SVM's.Source: Discussion · pp. 10–12
  • The authors speculated that CADx accuracy might be comparable to radiologists, but differences in CT images, label quality, and SVM kernel make precise comparisons difficult.Source: Discussion · p. 10
  • The authors cite XGBoost's invariance to feature-vector scaling as a prime reason for its superiority over scale-dependent SVM kernels.Source: Discussion · p. 11
  • In general, TPE performed better than random search for optimizing SVM and XGBoost parameters, but the performance gap was minimal at 10 trials and at 1000 trials for SVM.The authors surmised that SVM's narrower parameter space than XGBoost's let both methods almost fully optimize parameters, potentially explaining the minimal difference.Source: Discussion · p. 11
  • The small nodule sample could permit overfitting, so the authors call for larger samples to evaluate generalizability.Source: Discussion · p. 11
  • The study assessed technical usefulness, not clinical utility; it restricted feature-vector search after instability and omitted deep-learning CADx, which might outperform it.Source: Discussion · p. 12

Tables

The tables report CADx validation loss, AUC, and accuracy for SVM and XGBoost under random search and TPE across trial counts. XGBoost with TPE at 1,000 trials has the highest reported AUC in these tables.

  • 0.850 AUC was reported for SVM with TPE at 1,000 trials.Source: Tables · p. 19, Table 1
  • 0.896 AUC was reported for XGBoost with TPE at 1,000 trials.Source: Tables · p. 20, Table 2
  • At 1,000 trials, XGBoost with TPE had 0.394 validation loss and 0.820 accuracy.Source: Tables · p. 20, Table 2

Figure and Figure legends

The figures identify the CADx system outline and report validation loss, AUC, and accuracy. The legends also expand abbreviations used for the system and methods.

Related papers

Loading 1708.05897v2…