Source-linked AI summary
Chest X-Ray Analysis of Tuberculosis by Deep Learning with Segmentation and Augmentation
Sergii Stirenko, Yuriy Kochura, Oleg Alienin, Oleksandr Rokovyi, Peng Gang, Wei Zeng, Yuri Gordienko
TL;DR
The paper asks whether tuberculosis CADx can make reliable predictions from small, heterogeneous CXR datasets, where deep learning was thought to perform poorly. It applies lung segmentation and lossless or lossy augmentation to SH images; lossless augmentation yields the lowest validation loss without overfitting and nearly the same accuracy, while limited lossy augmentation lowers validation loss but reduces accuracy.
Problem
The paper addresses whether deep learning can provide reliable TB predictions from small, variable CXR datasets, despite concerns that such datasets yield poor accuracy.
Source: §I · p. 1, §II · p. 2Method
The study applies lung segmentation and lossless or lossy data augmentation to the Shenzhen Hospital CXR dataset and trains a deep CNN for tuberculosis diagnosis.
Source: §II · p. 2, §V · p. 6Results
Lossless augmentation of segmented data produced the lowest validation loss without overfitting and nearly the same accuracy as the original and lossy-augmented data; limited lossy augmentation lowered validation loss but reduced validation accuracy.
Source: §V · pp. 5–6Takeaways & Limitations
For small, imbalanced datasets, the authors identify segmentation, augmentation, better-balanced subsets, and exclusion of non-evident outliers as possible routes to improved CADx progress.
Source: §V · pp. 5–6The evaluation centers on the 662-image SH dataset, and the authors identify larger datasets as an improvement path.
Source: §V · p. 5
Abstract
from arXiv · showhide
The results of chest X-ray (CXR) analysis of 2D images to get the statistically reliable predictions (availability of tuberculosis) by computer-aided diagnosis (CADx) on the basis of deep learning are presented. They demonstrate the efficiency of lung segmentation, lossless and lossy data augmentation for CADx of tuberculosis by deep convolutional neural network (CNN) applied to the small and not well-balanced dataset even. CNN demonstrates ability to train (despite overfitting) on the pre-processed dataset obtained after lung segmentation in contrast to the original not-segmented dataset. Lossless data augmentation of the segmented dataset leads to the lowest validation loss (without overfitting) and nearly the same accuracy (within the limits of standard deviation) in comparison to the original and other pre-processed datasets after lossy data augmentation. The additional limited lossy data augmentation results in the lower validation loss, but with a decrease of the validation accuracy. In conclusion, besides the more complex deep CNNs and bigger datasets, the better progress of CADx for the small and not well-balanced datasets even could be obtained by better segmentation, data augmentation, dataset stratification, and exclusion of non-evident outliers.
I. INTRODUCTION
The introduction frames deep learning for chest X-ray diagnosis as promising, but notes expert concern that small datasets yield poor predictions. This work describes segmentation and augmentation to seek reliable tuberculosis predictions from fewer than 1,000 images.
I. INTRODUCTION
- Experts expected deep learning to produce very low accuracy on datasets with fewer than 1,000 images, despite progress in automated chest X-ray analysis.The introduction cites advances in GPU hardware, medical-image software, and deep learning, including disease detection at a level exceeding certified radiologists.Source: §I · p. 1
- The proposed approach combines lung segmentation with lossless and lossy data augmentation for small-dataset diagnosis.Source: §I · p. 1
- The paper targets statistically reliable tuberculosis-availability predictions from a dataset with fewer than 1,000 images.Source: §I · p. 1
II. PROBLEM AND RELATED WORK
Prior CXR datasets range from hundreds of images to a large public collection, while small-dataset diagnosis remains challenged by patient and image variability. The paper examines segmentation and augmentation for tuberculosis CADx on the small Shenzhen Hospital dataset.
II. PROBLEM AND RELATED WORK
- Small CXR datasets vary in image size, quality, patient characteristics, and artifacts, while regions outside the lungs can affect analysis.The cited public datasets include JSRT (247 images), LIDC (~1,000), Montgomery County (138), Shenzhen Hospital (662), and ChestX-ray14 (112,120 images).Source: §II · p. 1
- Prior CXR cancer work demonstrated dimensionality-reduction approaches including lung segmentation, bone-shadow exclusion, and t-SNE outlier exclusion.Source: §II · p. 1
- Despite such work, CADx remains challenging amid variability in lungs and surrounding regions; this study tests segmentation and lossless/lossy augmentation for TB prediction on SH-sized data.The stated aim is statistically reliable predictions on datasets like SH with fewer than 1,000 images.Source: §II · p. 2
III. EQUIPMENT, METHODS, AND DATA
The methods use a CNN trained on multiple GPUs and the Shenzhen Hospital dataset, whose demographic and image-area distributions are uneven. Manually prepared lung masks create a segmented dataset, building on prior evidence that external segmentation can aid small-dataset training.
III. EQUIPMENT, METHODS, AND DATA
- The CNN was trained on NVIDIA Tesla K40, GTX 1080 Ti, and Tesla V100 GPUs; Table I reports performance, bandwidth, core counts, training times, and speedups.The model was not optimized for V100 tensor processing units, so the authors say its speedup is not especially impressive.Source: §III · p. 2, Table I · p. 2
- The 662-image SH dataset contains normal and tuberculosis-marked abnormal CXRs and is not well balanced by disease status, age, or gender.Image dimensions vary, producing a wide, asymmetric distribution of image areas.Source: §III · p. 2
- Earlier attempts to train on small CXR datasets without preprocessing failed, while external lung-field segmentation had been reported to support successful training and improve prediction accuracy.Source: §III · p. 2
- Manually prepared left- and right-lung masks produce the segmented SH dataset; mask areas have a more nearly normal distribution than the asymmetric image-area distribution.The authors attribute mask variability partly to outliers, including children with smaller lungs.Source: §III · p. 2, Figure 3 · p. 2
- Figure 4 presents examples of the most dissimilar, most similar, and average lung masks.Source: Figure 4 · p. 2, §III · p. 2
IV. RESULTS
The results-section passage describes the dataset split and training comparisons, but gives no outcome values. It evaluates a CNN on segmented SH images with and without two forms of augmentation.
IV. RESULTS
- The CNN predicts tuberculosis presence in 336 images or absence in 326 images, using an 8:1:1 training, validation, and test split.Source: §IV · p. 2
- Images were rescaled by 1./255 and resized to 2048×2048 before assignment to the three dataset splits.Source: §IV · p. 2
- Training and validation runs of up to 100 epochs compared segmented SH data with versions using lossless or lossy augmentation.The CNN included dropout.Source: §IV · p. 2
A. Segmented Dataset
For the segmented CXR, training accuracy approached 1 while validation accuracy stayed at or below 0.76, with rising validation loss indicating overfitting.
A. Segmented Dataset
- Accuracy and loss curves were averaged across runs and then smoothed using locally weighted polynomial regression.The averaged curves are red and the smoothed curves blue.Source: A. Segmented Dataset · p. 3
- Validation accuracy stayed at or below 0.76 while validation loss grew, despite training accuracy approaching 1 and training loss tending to 0.The authors attribute this overfitting to the dataset's small size and investigate data augmentation.Source: A. Segmented Dataset · p. 3, Figure 5 · p. 3
- The observed overfitting motivated investigation of lossless and lossy data augmentation.Source: A. Segmented Dataset · p. 3
B. Effect of Lossless Data Augmentation
Lossless augmentation expanded the segmented dataset eightfold through reflections and 90-degree rotations. Its validation loss reached a global minimum corresponding to validation accuracy of 0.7±0.1.
B. Effect of Lossless Data Augmentation
- Reflections and 90-degree rotations expanded the segmented dataset by 8 times.The transformations included left-right and up-down reflections and rotations by 90n degrees, where n = 1,2,3.Source: B. Effect of Lossless Data Augmentation · p. 3
- Training accuracy did not exceed 0.9, validation accuracy did not exceed 0.7, and training loss declined toward 0 more slowly.Source: B. Effect of Lossless Data Augmentation · p. 3, Figure 7 · p. 3, Figure 6 · p. 3
- At the global validation-loss minimum, validation accuracy was 0.7±0.1.The minimum occurred at approximately 3000 steps and was described as the most realistic validation-accuracy estimate.Source: Figure 8 · p. 4, B. Effect of Lossless Data Augmentation · p. 4
C. Effect of Lossy Data Augmentation
Adding 5-degree rotations lowered validation accuracy relative to lossless augmentation, while validation loss remained lower than for the segmented dataset up to the non-overfitted local minimum.
C. Effect of Lossy Data Augmentation
- Adding 5-degree rotations produced training accuracy below 0.85, validation accuracy below 0.65, and training loss no lower than 0.35.Source: C. Effect of Lossy Data Augmentation · p. 4, Figure 9 · p. 4
- Augmented datasets had lower validation-loss minima than the segmented dataset, but that comparison held only up to the local minima associated with non-overfitted training.Validation-accuracy standard deviations were also lower for both augmented datasets despite fewer runs.Source: C. Effect of Lossy Data Augmentation · p. 4
- At the validation-loss global minimum, validation accuracy was 0.64±0.1.The minimum occurred at approximately 2500 steps.Source: Figure 10 · p. 4, C. Effect of Lossy Data Augmentation · p. 4
V. DISCUSSION AND CONCLUSIONS
The discussion finds that lung segmentation makes training on this small, imbalanced tuberculosis dataset possible, while lossless augmentation gives the strongest validation-loss outcome without overfitting. Limited lossy augmentation lowers validation loss but also lowers validation accuracy, and broader augmentation effects remain an open question.
V. DISCUSSION AND CONCLUSIONS
- Limited lossy augmentation lowered validation loss but also lowered validation accuracy, with training and validation accuracies remaining apart.The tested lossy augmentation used rotation up to 5 degrees; the authors link the accuracy trade-off to information loss.Source: §V · pp. 5–6
- Lung segmentation enabled CNN training on the small, imbalanced dataset, unlike the original unsegmented data, but training on segmented images alone overfit.The segmented-only run had higher training accuracy alongside rising validation loss and a widening training-validation gap.Source: §V · p. 5
- The effects of broader lossy augmentation remain open, including how different techniques affect CADx predictions across heterogeneous datasets.The authors call for further investigation of transformations such as greater rotation, scaling, and distortion.Source: §V · p. 5
- Beyond larger datasets and more complex CNNs, the authors propose better dataset stratification, lung and bone segmentation, and exclusion of non-evident outliers.Suggested stratification includes separating patient groups by characteristics such as age and gender.Source: §V · pp. 5–6
- Lossless augmentation of segmented images produced the lowest validation loss without overfitting and nearly the same accuracy as the original and lossy-augmented datasets.The comparison covers the averaged, smoothed accuracy and loss curves for the segmented, lossless-augmented, and lossy-augmented datasets.Source: §V · pp. 5–6, Figure 11 · p. 5