Source-linked AI summary
Automatic Lung Cancer Prediction from Chest X-ray Images Using Deep Learning Approach
Worawate Ausawalaithong, Sanparith Marukatat, Arjaree Thirach, Theerawit Wilaiprasitporn
TL;DR
Rural areas have limited access to LDCT and CT, while chest x-rays are more available but produce lower-quality images. This study uses DenseNet-121 with staged transfer learning from ImageNet to lung nodules and then lung cancer, achieving promising classification performance and generating nodule-localization heatmaps.
Problem
Limited rural access to LDCT and CT motivates improving lung cancer diagnosis from more readily available chest x-rays.
Method
DenseNet-121 was trained through staged transfer learning from ImageNet to chest-x-ray lung nodules and finally lung cancer to address the small dataset.
Results
The proposed strategy performed better than normal transfer learning, with higher mean accuracy and mean sensitivity.
Takeaways & Limitations
The model can classify lung cancer from chest x-rays and produce heatmaps showing the predicted position of lung cancer or nodules.
Abstract
from arXiv · showhide
Since, cancer is curable when diagnosed at an early stage, lung cancer screening plays an important role in preventive care. Although both low dose computed tomography (LDCT) and computed tomography (CT) scans provide more medical information than normal chest x-rays, there is very limited access to these technologies in rural areas. Recently, there is a trend in using computer-aided diagnosis (CADx) to assist in screening and diagnosing of cancer from biomedical images. In this study, the 121-layer convolutional neural network also known as DenseNet-121 by G. Huang et. al., along with the transfer learning scheme was explored as a means to classify lung cancer using chest X-ray images. The model was trained on a lung nodules dataset before training on the lung cancer dataset to alleviate the problem of a small dataset. The proposed model yields 74.43$\pm$6.01\% of mean accuracy, 74.96$\pm$9.85\% of mean specificity, and 74.68$\pm$15.33\% of mean sensitivity. The proposed model also provides a heatmap for identifying the location of the lung nodule. These findings are promising for further development of chest x-ray-based lung cancer diagnosis using the deep learning approach. Moreover, these findings solve the problem of small dataset.
I. INTRODUCTION
The study addresses limited rural access to LDCT and CT by exploring chest X-rays with CADx for lung cancer diagnosis. It also targets small datasets through repeated transfer learning and adds heatmaps to indicate predicted nodule locations.
- Motivation: LDCT screening reduces lung cancer deaths but faces equipment, financial, and specialist-access barriers in rural populations.Chest X-rays are more readily available, though their lower image quality generally limits diagnostic quality.
- Motivation: Chest X-rays are evaluated with a CADx system to improve lung cancer diagnostic performance despite lower image quality than LDCT or CT.
- Research gap: Small lung cancer image datasets make conventional CNN transfer learning difficult because CNNs contain many parameters.The proposed approach applies transfer learning repeatedly to alleviate the small-sample problem.
- Contribution: Multi-transfer learning is reported to achieve better task performance than traditional transfer learning.
- Contribution: The model can produce a heatmap showing the predicted position of lung cancer in a chest X-ray image.
II. METHODOLOGY
The methodology section presents the study workflow, moving from dataset descriptions and preparation to model architecture, loss function, and visualization.
- II. METHODOLOGY: The study describes datasets, data preparation, model architecture, loss function, and visualization in sequence.
A. Datasets
The JSRT dataset contains 247 frontal chest X-ray images, including malignant, benign, and non-nodule cases, and is evaluated with 10-fold cross-validation.
- A. Datasets: The JSRT dataset contains 247 frontal chest X-ray images, including 154 with lung nodules and 93 without nodules.The nodule-positive images comprise 100 malignant and 54 benign cases.
- A. Datasets: JSRT images are 2048 × 2048 pixels and are evaluated using 10-fold cross-validation because the dataset is very small.
2) ChestX-ray14 Dataset [17]:
ChestX-ray14 is a large public chest X-ray dataset with 112,120 frontal images and up to 14 thoracic pathology labels per image, but it contains no lung cancer images.
- 2) ChestX-ray14 Dataset [17]: ChestX-ray14 contains 112,120 frontal chest X-ray images, making it the largest public chest X-ray dataset described here.
- 2) ChestX-ray14 Dataset [17]: Each ChestX-ray14 image has up to 14 thoracic pathology labels and a size of 1024 × 1024 pixels.
- 2) ChestX-ray14 Dataset [17]: ChestX-ray14 does not contain lung cancer images.
B. Data Preparation
The data preparation pipeline applies four preprocessing steps to standardize chest X-ray images before model input: contrast enhancement, denoising, resizing, and color normalization.
- Data preparation consists of four preprocessing steps applied to all images.The process is illustrated using an example image from the JSRT dataset.
- Histogram Equalization increases image contrast and normalizes intensity across datasets.
- Median Filtering removes image noise using a 3 × 3 window.
- Images are resized to 224 × 224 pixels to match the model input.
- Image color is normalized using the ImageNet training set mean and standard deviation.
C. CNN architecture and transfer learning
The study uses DenseNet-121 with staged transfer learning to adapt from general images to chest X-rays, nodules, and ultimately malignant-nodule classification.
- CNN architecture: DenseNet-121 replaces its original 1000-category output layer with a single sigmoid node for pathology probability.The architecture uses dense blocks designed to improve data flow and address vanishing gradients.
- Transfer learning: Transfer learning is applied twice: first for nodule classification, then for malignant-nodule classification.The second task treats benign-nodule and non-nodule images as the non-malignant class.
- Training sequence: The Base Model starts from DenseNet-121 weights pretrained on ImageNet with a one-class sigmoid output.
- Training sequence: Retrained Model A adapts the Base Model on ChestX-ray14 using nodules as positive cases and non-nodules as negative cases.Training images were randomly flipped horizontally.
- Training sequence: Retrained Model B trains on JSRT for malignant versus non-malignant nodule classification using augmentation and 10-fold cross-validation.Training images were randomly rotated within 30 degrees and randomly flipped horizontally.
- Training sequence: Retrained Model C further retrains Model A on JSRT to adapt chest X-ray and nodule knowledge to malignant-nodule identification.The training sequence progresses from ImageNet to ChestX-ray14 and then JSRT; Model C showed the best performance compared with Model B.
D. Loss and optimizer
The study uses a weighted binary classification loss for imbalanced transfer-learning datasets and optimizes the model with Adam using a decaying learning rate.
- A weighted binary classification loss is used because both transfer-learning datasets are imbalanced.
- The loss uses image labels and predicted class probabilities, with weights based on the proportions of negative and positive cases.The weights are intended to make training more efficient.
- Adam optimization uses β1 = 0.9, β2 = 0.999, an initial learning rate of 0.001, and batch size 32.The learning rate decreases by a factor of 10 when validation loss plateaus.
E. Class Activation Mappings (CAMs)
Class Activation Mappings derived from Model C identify the image regions most salient to the model’s predicted class.
- Class Activation Mappings show the most salient image locations used by Model C to identify the output class.
- The CAM map combines feature maps using their final-layer weights for the selected class.Mc denotes the salient-feature map, wc,k the final-layer weight, and fk the kth feature map.
III. RESULTS AND DISCUSSION
The study evaluates retrained models for lung-nodule and lung-cancer classification, comparing their diagnostic metrics and visual localization of malignant regions. Retrained Model C generally outperforms Model B and produces mostly accurate Class Activation Maps, although dataset-size-related overfitting remains.
- Evaluation: Accuracy, specificity, and sensitivity were used to evaluate lung-nodule and lung-cancer classification performance.Model A was tested on ChestX-ray14, while Models B and C were evaluated on JSRT using 10-fold cross-validation.
- Model comparison: Model B achieved higher specificity, but Model C achieved better accuracy and sensitivity with lower standard deviations across all metrics.The comparison supports retraining for task-specific classification across the evaluated metrics.
- Model comparison: Retraining the model for specific tasks produced better results in almost all evaluation metrics.This finding is reported as a comparison between the retrained models’ performances.
- Localization: Model C produced accurate Class Activation Maps for most correctly predicted images, but some maps covered overly large or inaccurate regions.The blue circle marks the actual lung-cancer location in the JSRT images.
- Localization: Model C’s localization errors were attributed to slight overfitting caused by the small dataset, while Model B’s maps were generally less accurate.Model B often used an overly large image area to classify lung cancer.
IV. CONCLUSION
The conclusion presents a stepwise DenseNet-121 training strategy for lung-cancer detection from chest X-rays when the available dataset is small. It reports improved performance over normal transfer learning and mostly accurate cancer localization, while identifying overfitting and several future enhancements.
- Conclusion: The proposed strategy trains the network step-by-step from ImageNet to ChestX-ray14 nodules and finally JSRT lung cancer.This sequence is intended to train the model using a very small dataset.
- Conclusion: The proposed training strategy achieved higher mean accuracy and sensitivity but lower mean specificity, together with lower standard deviation than normal transfer learning.These comparisons are reported for the proposed method versus normal transfer learning.
- Conclusion: Model C produced quite accurate Class Activation Map locations for lung cancer, although some maps were affected by overfitting.Retrained Model B produced much poorer Class Activation Maps.
- Future work: Future work includes Gaussian-noise addition, random cropping, attention-guided nodule cropping, clinical feature integration, and ensemble modeling.The proposed additions are intended to support malignancy identification or improve results.