Source-linked AI summary
Machine Learning-based Lung and Colon Cancer Detection using Deep Feature Extraction and Ensemble Learning
Md. Alamin Talukder, Md. Manowarul Islam, Md Ashraf Uddin, Arnisha Akhter, Khondokar Fida Hasan, Mohammad Ali Moni
TL;DR
Lung and colon cancer histopathology requires efficient detection because these cancers are major causes of death and large clinical datasets are difficult to analyze traditionally. The paper proposes a hybrid model combining deep feature extraction, high-performance filtering, and ensemble learning on LC25000 images. It reports 99.05% accuracy for lung cancer, 100% for colon cancer, and 99.30% for combined lung-and-colon cancer detection.
Problem
Large healthcare datasets are difficult to analyze with traditional approaches, while lung and colon cancers are major causes of death.
Method
The paper evaluates a hybrid framework combining preprocessing, k-fold cross-validation, deep feature extraction, high-performance filtering, and ensemble learning on LC25000 histopathological images.
Results
The model achieved 99.05% accuracy for lung cancer, 100% for colon cancer, and 99.30% for combined lung-and-colon cancer detection.
Takeaways & Limitations
The authors conclude that the proposed models could support doctors in cancer diagnosis and outperform other techniques.
Takeaways & Limitations
The model depends on clear and sharp images for feature extraction, which the authors identify as an inadequacy.
Abstract
from arXiv · showhide
Cancer is a fatal disease caused by a combination of genetic diseases and a variety of biochemical abnormalities. Lung and colon cancer have emerged as two of the leading causes of death and disability in humans. The histopathological detection of such malignancies is usually the most important component in determining the best course of action. Early detection of the ailment on either front considerably decreases the likelihood of mortality. Machine learning and deep learning techniques can be utilized to speed up such cancer detection, allowing researchers to study a large number of patients in a much shorter amount of time and at a lower cost. In this research work, we introduced a hybrid ensemble feature extraction model to efficiently identify lung and colon cancer. It integrates deep feature extraction and ensemble learning with high-performance filtering for cancer image datasets. The model is evaluated on histopathological (LC25000) lung and colon datasets. According to the study findings, our hybrid model can detect lung, colon, and (lung and colon) cancer with accuracy rates of 99.05%, 100%, and 99.30%, respectively. The study's findings show that our proposed strategy outperforms existing models significantly. Thus, these models could be applicable in clinics to support the doctor in the diagnosis of cancers.
1. Introduction
The introduction frames lung and colon cancer detection as an important clinical problem and motivates a hybrid machine-learning framework for histopathological images. The proposed approach combines preprocessing, cross-validation, deep feature extraction, high-performance filtering, and ensemble learning, achieving accuracy above 99% across the reported cancer tasks.
- Clinical motivation: Lung and colorectal cancers are presented as major causes of death, with reported mortality proportions of 12.2% and 10.7%, respectively.The introduction identifies lung cancer as a commonly recognized malignancy and colorectal cancer as affecting the colon and rectum.
- Motivation: Large healthcare datasets are difficult to analyze and extrapolate with traditional approaches, motivating machine-learning and deep-learning methods.The paper links growing data archives and increased healthcare-data availability to opportunities for more comprehensive clinical analysis.
- Technical background: Transfer learning can either train image datasets as a baseline or extract image features for subsequent machine- or deep-learning classification.The introduction describes transfer learning as repurposing acquired knowledge for a specific situation and distinguishes these two uses.
- Proposed framework: The proposed framework applies preprocessing, k-fold cross-validation, feature extraction, high-performance filtering, and ensemble learning to LC25000 histopathological images.The framework is designed to include cancer possibilities in a large dataset and uses an ensemble method at the end of the implementation.
- Evaluation: The study evaluates accuracy, recall, precision, F1-score, MAE, MSE, RMSE, confusion matrices, AUC, and ROC curves.These indicators are used to assess the proposed framework experimentally.
- Reported performance: 99%+ accuracy is reported for the combined lung-and-colon cancer classification task.The introduction states that the classifier algorithms detect combined lung and colon cancer with an accuracy rate above 99%.
2. Literature Review
The literature review surveys machine-learning, deep-learning, transfer-learning, feature-fusion, and ensemble approaches for lung and colon cancer histopathology. Reported accuracies range from 91% to 100%, while several studies indicate benefits from combining complementary features or models.
- Existing approaches: Deep-learning and transfer-learning methods have been widely applied to lung, colon, and other cancer histopathology images.The reviewed studies use CNNs, pretrained models, image-processing methods, and machine-learning classifiers across multiple datasets.
- CNN studies: 99.67% accuracy and a 1.24% loss rate were reported by MobileNetV2, exceeding the other two CNN models evaluated for colon cell imaging.Average pooling and max-pooling models achieved 95.48% and 97.49% accuracy, respectively.
- Feature fusion: 96.33% accuracy was reported for a CNN trained on combined 2D Fourier and 2D Wavelet features for cancer-tissue identification.The approach combined the extracted features before CNN training.
- Deep feature extraction: DenseNet121 features paired with KNN achieved 98.53% accuracy, outperforming ResNet50 features across the evaluated machine-learning algorithms.This study evaluated deep feature extraction for colorectal cancer diagnosis using two transfer-learning models and traditional classifiers.
- Combined features: A cooperative-feature Random Forest achieved 93.73% accuracy, supporting the use of mixed deep and handcrafted features rather than a single extractor.The reviewed study combined VGG19 features with SIFT, SURF, ORB, and Shi-Tomasi descriptors before classification.
- Ensemble methods: An ensemble of three CNN models achieved 91% validation accuracy and exceeded the individual models in the reviewed lung-cancer study.The individual models included VGG16, ResNet50V2, and DenseNet201.
3. Methodology
The methodology combines image preprocessing, transfer-learning feature extraction, machine-learning evaluation, high-performance filtering, and ensemble voting for LC25000 lung and colon cancer images.
- Overall framework: The proposed model applies preprocessing, deep feature extraction, high-performance filtering, and ensemble learning to lung and colon cancer images.The workflow is developed using LC25000 histopathological datasets.
- Dataset: The combined dataset includes lung adenocarcinoma, squamous cell carcinoma, benign tissue, colon adenocarcinoma, and colon benign tissue classes.The lung dataset has three cell labels, while the colon dataset has two.
- Pre-processing: Images are resized to 128x128, converted from BGR to RGB, represented as NumPy arrays, scaled by dividing pixel values by 255, and labeled.The scaling normalizes image intensity values to a range from 0 to 1.
- Filtering and ensemble learning: High-performance filtering selects the top three algorithms by accuracy for subsequent ensemble learning.Both majority hard voting and weighted-average soft voting are applied, and the best ensemble is selected for each transfer-learning model.
- Feature extraction and classification: Features are extracted with transfer-learning models including VGG16, VGG19, DenseNet169, and DenseNet201 before machine-learning classification.The extracted features are evaluated with Random Forest, SVM, Logistic Regression, MLP, XGB, and LGB.
- Evaluation: SVM, Logistic Regression, and MLP outperform the other evaluated machine-learning algorithms for lung, colon, and combined cancer graphs.The study evaluates selected models using accuracy, recall, precision, F1-score, error metrics, confusion matrices, AUC, and ROC curves.
4. Experimental setup and performance metrics
The experiments evaluate the cancer-detection model using a Windows 10 Pro machine, common Python libraries, 10-fold cross-validation, and multiple classification and error metrics. These metrics include confusion-matrix measures, ROC/AUC analysis, and MAE, MSE, and RMSE.
- Experimental environment: The experiments use Microsoft Windows 10 Pro with an Intel Core i3-6006U CPU, 8GB RAM, and 120GB SSD plus 1TB HDD storage.Implementation uses Anaconda Navigator and a Jupyter notebook.
- Experimental environment: Python libraries including pandas, NumPy, Matplotlib, Seaborn, TensorFlow, Keras, and Scikit-learn implement the suggested model.
- Performance metrics: Performance is assessed with accuracy, recall, precision, f1-score, ROC curve, confusion matrix, MSE, MAE, and RMSE.The confusion matrix organizes four predicted-versus-actual combinations: TP, TN, FP, and FN.
- Classification metrics: Accuracy measures correctly predicted observations relative to total observations, while precision and recall quantify positive prediction correctness and recovery of actual positives.The reported accuracy formulation uses TP, TN, FP, and FN.
- ROC and AUC: ROC analysis examines classifier effectiveness across thresholds using true-positive rate and false-positive rate, while AUC reflects class separability.An AUC approaching 1 indicates strong separability according to the supplied description.
- Validation procedure: 10-fold cross-validation partitions the dataset into 80% training and 20% testing, using each of 10 folds as a testing component.The procedure averages values computed across the folds as an evaluation metric.
5. Performance Evaluation and Analysis
The evaluation compares transfer-learning backbones and hard versus soft voting across lung, colon, and combined cancer datasets. MobileNet with soft voting is selected for the proposed model, achieving high reported accuracy and AUC values.
- Transfer-learning and voting comparison: Soft voting outperformed hard voting for the evaluated transfer-learning models and was selected for lung and colon cancer datasets.For lung cancer, average accuracy was 97.66% for hard voting and 98.05% for soft voting; for colon cancer, both were 99.57%.
- Lung cancer evaluation: 99.05% accuracy was reported for lung cancer using the selected MobileNet soft-voting configuration.The reported lung metrics were 99.05% accuracy, 99.03% precision, 99.06% recall, and 99.04% F1-score.
- Lung cancer evaluation: 99.96% AUC was reported for the lung cancer soft-voting classifier, with high true-positive and true-negative rates and low false-positive and false-negative rates.The confusion-matrix rates are reported separately for lung squamous cell carcinoma, lung adenocarcinomas, and lung benign tissue.
- Colon cancer evaluation: 100% accuracy was reported for colon cancer using MobileNet with soft voting.Accuracy, precision, recall, and F1-score were each 100%, while MAE, MSE, and RMSE were each 0%.
- Colon cancer evaluation: 100% AUC was reported for the colon cancer soft-voting classifier, with 44.29% true-positive, 55.71% true-negative, and zero false-positive and false-negative rates.The evaluation covered colon adenocarcinomas and colon benign tissue.
- Discussion: The proposed hybrid model is reported to achieve the highest accuracy among the compared methods on the evaluated datasets.The comparison used the LC25000 dataset, with slightly different image counts across methods.
6. Conclusion
The paper presents a hybrid lung and colon cancer detection model combining preprocessing, transfer-learning feature extraction, machine-learning algorithms, and ensemble voting. Evaluated on LC25000 histopathological datasets, it reports high accuracy while identifying image preprocessing and clarity as areas for further improvement.
- 6. Conclusion: The hybrid pipeline combines preprocessing, 10-fold cross-validation, transfer-learning feature extraction, machine-learning algorithms, and ensemble learning.The tested transfer-learning models included VGG16, VGG19, MobileNet, DenseNet169, and DenseNet201; evaluated machine-learning algorithms included RF, SVM, LR, MLP, XGB, and LGB.
- 6. Conclusion: MobileNet was selected for feature extraction, while an ensemble soft voting classifier was selected from two ensemble models.
- 6. Conclusion: The experiments used LC25000 lung and colon histopathological image datasets and evaluated performance with multiple metrics.The reported metrics included accuracy, recall, precision, F1-score, ROC curve, MAE, MSE, and RMSE.
- 6. Conclusion: 99.05% accuracy was reported for lung cancer detection, 100% for colon cancer, and 99.30% for combined lung and colon cancer detection.
- 6. Conclusion: The authors state that the developed model could support automated clinical diagnosis of lung and colon cancers.
- 6. Conclusion: Further preprocessing and model development are needed because effective feature extraction depends on clear and sharp images.The paper identifies image clarity and sharpness as an inadequacy and proposes exploring newer datasets with efficient preprocessing.