Source-linked AI summary
Deep Learning with Lung Segmentation and Bone Shadow Exclusion Techniques for Chest X-Ray Analysis of Lung Cancer
Yu. Gordienko, Peng Gang, Jiang Hui, Wei Zeng, Yu. Kochura, O. Alienin, O. Rokovyi, S. Stirenko
TL;DR
The paper addresses automated lung-cancer lesion and nodule analysis from widely available chest X-rays, where interpretation is difficult and radiologist shortages constrain screening. It compares deep-learning training after lung segmentation and bone-shadow exclusion across four JSRT-derived datasets, finding that preprocessing supports training and that bone-shadow exclusion performs best among the tested preprocessed configurations.
Problem
Chest X-rays are widely available but difficult to interpret for lung disease, creating a need for automated assistance amid limited radiologist capacity.
Method
The study trains a simple seven-layer CNN on original, bone-shadow-excluded, segmented, and combined-preprocessing versions of the JSRT dataset.
Results
Preprocessed datasets show a tendency to train, while bone-shadow-excluded BSE-JSRT (#02) achieves much better accuracy and loss results than the other segmented preprocessed datasets.
Takeaways & Limitations
Bone-shadow elimination and segmentation are useful preprocessing techniques for training in the tested simplified configuration.
Takeaways & Limitations
The authors identify overtraining related to artifacts such as lung shape and border patterns, and propose larger images, more data, augmentation, and more complex models for improvement.
Abstract
from arXiv · showhide
The recent progress of computing, machine learning, and especially deep learning, for image recognition brings a meaningful effect for automatic detection of various diseases from chest X-ray images (CXRs). Here efficiency of lung segmentation and bone shadow exclusion techniques is demonstrated for analysis of 2D CXRs by deep learning approach to help radiologists identify suspicious lesions and nodules in lung cancer patients. Training and validation was performed on the original JSRT dataset (dataset #01), BSE-JSRT dataset, i.e. the same JSRT dataset, but without clavicle and rib shadows (dataset #02), original JSRT dataset after segmentation (dataset #03), and BSE-JSRT dataset after segmentation (dataset #04). The results demonstrate the high efficiency and usefulness of the considered pre-processing techniques in the simplified configuration even. The pre-processed dataset without bones (dataset #02) demonstrates the much better accuracy and loss results in comparison to the other pre-processed datasets after lung segmentation (datasets #02 and #03).
1 Introduction
Lung cancer creates a substantial screening burden, while CXRs offer broad availability but require difficult expert interpretation. Deep learning is presented as a potential source of automated assistance for this setting.
- Lung cancer has high incidence worldwide and in China, where heavy air pollution aggravates its consequences.
- Early screening can produce good outcomes at relatively low cost, but CT screening remains costly and limited in availability across much of the world.
- CXRs are widely available for diagnosing lung diseases, including cancer, but detecting disease marks is complicated and requires expert radiologists.
- Long manual analysis and shortages of radiologists limit lung-cancer detection during large-scale screening.
- Recent advances in GPUs, machine learning, and deep learning motivate automated assistance tools for medical image recognition.
2 Problem and Related Work
The related work positions digital technologies and machine learning as renewing the diagnostic role of X-rays. It also identifies lung-field segmentation as a preprocessing strategy for focusing predictions on relevant anatomy.
- Digital technologies and machine learning are described as reviving the significance of X-ray imaging for medical diagnosis.
- Machine-learning approaches can detect cardiothoracic lesions and are especially sensitive to lung nodules on X-ray scans.
- ChestX-ray14 contains over 100 000 frontal-view X-ray images covering 14 lung diseases, and CheXNet is a 121-layer CNN trained on it.
- Researchers seek higher prediction accuracy by excluding regions not pertinent to the lungs or other regions of interest.
- Lung-field segmentation methods have been tested on JSRT, where lung structures were manually segmented by two human observers.
3 Data and Methods
The study compares deep-learning training across original, bone-shadow-excluded, segmented, and combined-preprocessing versions of JSRT. It uses a simple CNN and evaluates computational behavior across CPU and GPU configurations.
- Data: JSRT contains 247 images: 154 cases with lung nodules and 93 cases without lung nodules.
- Data: BSE-JSRT contains the same 247 JSRT images after clavicle and rib shadows were removed by specialized algorithms.
- Lung segmentation: A UNet-based CNN predicts black-and-white lung masks, which are used to cut the left and right lung fields from surrounding anatomy.
- Model and computation: The study uses a simple seven-layer 2D CNN to isolate the effects of segmentation and bone elimination rather than maximize accuracy or minimize loss.
- Model and computation: Running time and speedup are evaluated for single-CPU, multi-CPU, and GPU modes across varying image and batch sizes.
4 Results
Training compares four dataset configurations: original JSRT, bone-shadow-excluded BSE-JSRT, segmented JSRT, and segmented BSE-JSRT. Preprocessing enables training in the simplified configuration, with bone exclusion producing the strongest reported accuracy and loss results among the preprocessed datasets.
- Experimental comparison: The experiment compares original JSRT (#01), BSE-JSRT without clavicle and rib shadows (#02), segmented JSRT (#03), and segmented BSE-JSRT (#04).
- Preprocessing: The segmentation pipeline trains a UNet-based CNN on MC masks, predicts lung borders, and cuts the right and left lung regions.
- Training and validation: All preprocessed datasets (#02, #03, and #04) show a tendency to train, whereas raw JSRT #01 shows no sign of training in the simplified configuration.
- Training and validation: Preprocessed dataset #02 achieves much better accuracy and loss results than the other preprocessed datasets after lung segmentation.
- Training and validation: Training accuracy and loss outperform validation accuracy and loss, indicating overtraining associated with artifacts such as lung shape and border patterns.
5 Discussion and Conclusions
Pre-processing with bone shadow elimination and lung segmentation was useful even in the simplified configuration. The bone-free dataset showed better accuracy and loss results than the other segmented pre-processed datasets, while overtraining remained associated with lung-shape and border artifacts.
- Bone shadow elimination and lung segmentation demonstrated usefulness for training in the simplified configuration.
- Training and validation curves showed overtraining, with lower validation accuracy and higher validation loss than training results.
- The overtraining effect was considered related to learning artifacts such as lung shape and lung-border patterns.
- Further improvement was proposed through larger images, more images, data augmentation, and integration of datasets from hospitals worldwide.
- The bone-free dataset achieved much better accuracy and loss results than the other pre-processed datasets after lung segmentation.
- Additional development was proposed for lung segmentation using larger datasets with masks and for more complex bone-shadow elimination techniques.