A CNN-based methodology for breast cancer diagnosis using thermal images

Juan Zuluaga-Gomez; Zeina Al Masry; Khaled Benaggoune; Safa Meraghni; Noureddine Zerhouni

Source-linked AI summary

A CNN-based methodology for breast cancer diagnosis using thermal images

Juan Zuluaga-Gomez, Zeina Al Masry, Khaled Benaggoune, Safa Meraghni, Noureddine Zerhouni

arXiv:1910.13757v1Oct 30, 2019cs.CVeess.IV

TL;DR

Breast thermography needs accessible, reliable computer-aided diagnosis, while limited public data and inconsistent evaluation frameworks constrain comparison. This study evaluates CNN-based diagnosis with controlled database splitting, augmentation, and hyper-parameter optimization, achieving strong performance and quantifying the trade-off between augmentation and database size.

Problem
Breast cancer diagnosis needs high-quality, cheap, and reliable computer-aided systems, but limited thermography databases and inconsistent training frameworks constrain evaluation and comparison.
Method
The study evaluates CNN architectures using a defined 50/20/30 train-validation-test split, data preprocessing and augmentation, and Bayesian hyper-parameter optimization with a tree parzen estimator.
Results
92% accuracy and 92% F1-score were obtained among 57 patients, while augmentation matched the performance of a database up to 50% larger.
Takeaways & Limitations
The study highlights data augmentation and CNNs for thermal breast images and provides baseline information for future thermography databases and CNN research.
Takeaways & Limitations
Limited thermal-image information constrains model generalization, while CNNs remain black-box models whose computations physicians and researchers want to understand.

Abstract

from arXiv · show

Micro Abstract: A recent study from GLOBOCAN disclosed that during 2018 two million women worldwide had been diagnosed from breast cancer. This study presents a computer-aided diagnosis system based on convolutional neural networks as an alternative diagnosis methodology for breast cancer diagnosis with thermal images. Experimental results showed that lower false-positives and false-negatives classification rates are obtained when data pre-processing and data augmentation techniques are implemented in these thermal images. Background: There are many types of breast cancer screening techniques such as, mammography, magnetic resonance imaging, ultrasound and blood sample tests, which require either, expensive devices or personal qualified. Currently, some countries still lack access to these main screening techniques due to economic, social or cultural issues. The objective of this study is to demonstrate that computer-aided diagnosis(CAD) systems based on convolutional neural networks (CNN) are faster, reliable and robust than other techniques. Methods: We performed a study of the influence of data pre-processing, data augmentation and database size versus a proposed set of CNN models. Furthermore, we developed a CNN hyper-parameters fine-tuning optimization algorithm using a tree parzen estimator. Results: Among the 57 patients database, our CNN models obtained a higher accuracy (92\%) and F1-score (92\%) that outperforms several state-of-the-art architectures such as ResNet50, SeResNet50 and Inception. Also, we demonstrated that a CNN model that implements data-augmentation techniques reach identical performance metrics in comparison with a CNN that uses a database up to 50\% bigger. Conclusion: This study highlights the benefits of data augmentation and CNNs in thermal breast images. Also, it measures the influence of the database size in the performance of CNNs.

1 Introduction

The paper motivates CNN-based computer-aided diagnosis with thermal images as an accessible breast-cancer screening approach and investigates preprocessing, augmentation, architecture, optimization, and database-size effects.

Motivation: Thermography can detect temperature changes linked to blood perfusion, inflammation, angiogenesis, and breast tumors.It may also address some mammography limitations, including x-rays, painfulness, and cost barriers.
Research gap: CNNs have shown promising medical-imaging results but have been used only sparsely for breast-cancer thermography in the DMR-IR database.The paper identifies possible concerns about efficiency relative to texture or statistical features and computational load.
Approach: The study proposes a CNN-CAD methodology for breast thermography and a benchmark against ResNet, SeResNet, VGG16, Inception, InceptionResNetV2, and Xception.It also introduces a tree-parzen-estimator hyper-parameter optimization algorithm.
Approach: The authors develop an unbiased training methodology intended to reduce likely overfitting in models trained on the DMR-IR database.Baseline CNN models replicate recent studies before addressing the training weakness.
Experimental scope: The study measures the trade-off between data augmentation and database size because biomedical databases are often small, expensive to acquire, and protocol-dependent.The stated goal is to suggest a minimum patient count for an effective CAD system.

2 Current techniques for breast cancer diagnosis from thermal images

Prior thermal-breast-cancer CAD studies largely use the DMR-IR database and texture or statistical features, while CNN studies form a smaller branch. The paper addresses comparability and complexity by defining a detailed CNN methodology and benchmarking alternatives.

Database and prior work: The literature includes machine-learning, fuzzy-logic, clustering, support-vector-machine, texture, and statistical-feature approaches for thermal breast-cancer diagnosis.One cited support-vector-machine study reported 88.1% accuracy, 85.71% sensitivity, and 90.48% specificity.
Database and prior work: The DMR-IR database contains 1140 images from 57 Brazilian patients, including 38 with anomalies and 19 healthy women, with 20 images per patient.It provides heat-map images and temperature matrices for breast-thermography research.
Database and prior work: Earlier DMR-IR studies use texture and statistical features or CNNs, with reported accuracies reaching 95% and exceeding 90%, respectively.Comparisons are complicated by differing patient counts and experimental setups.
Preprocessing: Thermal images contain irrelevant regions such as backgrounds and other body parts, motivating reduction of non-relevant information before classification.Reviews identify this reduction as an important need for CAD thermography systems.
Paper positioning: The paper argues that algorithmic complexity is not directly proportional to performance and targets CNN systems that are less complex, easier to train, and more generalizable to new patients.Its methodology also examines preprocessing, database splitting, augmentation, and database size.

3 Database description and proposed methodology

The methodology preprocesses and augments thermal breast images, evaluates baseline and benchmark CNNs, and uses Bayesian hyper-parameter optimization before computing diagnostic metrics.

Workflow: The pipeline acquires thermal data, preprocesses and augments images, trains CNN variants, and evaluates them with multiple performance metrics.Its workflow has three phases: data preparation, baseline/benchmark/fine-tuned modeling, and metric evaluation.
Database: The DMR-IR database contains 57 patients with diagnoses confirmed by mammography, ultrasound, and biopsies.It includes 19 healthy patients and 37 patients with malignant breast disease.
Pre-processing and augmentation: Preprocessing segments breast regions, crops, resizes, and normalizes thermal images to 250x300 temperature points, reducing computational cost by one quarter.Augmentation uses flips, 0-45° rotations, 20% zoom, and normalized noise.
CNN evaluation: The training framework uses a 50/20/30 train-validation-test split and compares baseline CNNs with ResNet, SeResNet, VGG16, Inception, InceptionResNetV2, and Xception.The stated split is intended to provide a detailed methodology for reducing bias and overfitting.
Hyper-parameter optimization: Bayesian optimization with a tree parzen estimator searches CNN hyper-parameters using a probabilistic surrogate model updated from prior results.The objective function trains a CNN, while evaluation returns accuracy, precision, sensitivity, F1-score, and ROC-AUC.
Evaluation metrics: F1-score is selected because it accounts for false positives and false negatives, whereas accuracy accounts for true positives and true negatives.The paper emphasizes these error types for breast cancer diagnosis and class imbalance.

4 Experimental setup up and results

Four experiments assess CNN training splits, benchmark architectures, Bayesian architecture optimization, and the effects of augmentation and database size. Results show that split methodology and model complexity strongly affect performance, while the optimized model achieves balanced diagnostic metrics.

Experimental setup: The study conducts four experiments covering baseline CNNs, state-of-the-art benchmarking, Bayesian architecture optimization, and training scenarios varying augmentation and database split.The fourth experiment specifically compares training with and without augmentation while varying database split ratios.
Baseline results: 99% accuracy and 99% F1 Score were obtained under Approach 1, compared with 88% accuracy and 89% F1-score under Approach 2 for CNN 1.The paper attributes the stronger Approach 1 results to possible patient-image overlap between training and test data.
Benchmark results: 30% improvement of F1-score was observed for Inception V3 when using global average pooling rather than a flatten layer.Global average pooling and Adam were subsequently used for the state-of-the-art CNN models.
Benchmark results: More complex CNN architectures performed worse on the small DMR-IR database than the authors’ simpler architectures.The benchmark models were designed for multi-class classification on large ImageNet-like datasets, unlike this binary task.
Optimization results: 92% accuracy, 98% precision, 87% sensitivity, 92% F1-score, and 92% ROU-AUC were achieved by the optimized top model.The best configuration used GAP or flatten layers, dropout between 0 and 0.3, and 6 or 7 CNN blocks.
Interpretation: The authors identify bias and overfitting in Approach 1 and use subsequent benchmark and optimization experiments to address weaknesses in model robustness.Figure 6 summarizes averaged metrics from experiments 1 to 3 in ascending experimental order.

5 Discussion and conclusion

The study develops and evaluates CNN-based computer-aided diagnosis for breast thermography, emphasizing unbiased database splitting, architecture optimization, and the trade-off between data augmentation and dataset size. Results support CNN-based thermal-image systems as potentially useful additional tools while identifying limited data, interpretability, and output-format constraints.

Study scope: The study compares CNN performance across database-splitting methodologies, benchmarks state-of-the-art architectures, and evaluates hyperparameter optimization, data augmentation, and dataset size.These experiments provide baseline information for future breast-thermography databases and studies.
Evaluation methodology: Patient-level database splitting reduced bias and overfitting compared with splits that could place images or sequences from one patient across datasets.The robust framework assigns all images or sequences from each patient to training, testing, or validation, whereas the alternative produced training metrics above 98% for the top model.
CNN optimization: Bayesian optimization produced CNN architectures that overcame previous models while combining low variance, low bias, and low overfitting in the reported experiments.Experiment 3 used a tree-parzen-estimator-based search rather than selecting architectures empirically.
Data requirements: Data augmentation generally outperformed no augmentation, while increasing the database from 10 to 47 patients improved generalization and performance.Performance variance decreased as the database grew, with stabilization observed between 40 and 47 patients.
Limitations: The study identifies limited thermal-image availability, CNN black-box behavior, and the need to provide both an image and cancer probability as remaining limitations.These constraints limit generalization, interpretability, and alignment with physicians’ preferred system outputs.
CNN optimization: 92% accuracy, 94% precision, 91% sensitivity, and 92% F1-score were achieved by the top Bayesian-optimized CNN model.The model was evaluated in the DMR-IR thermography database.