Derin öğrenme teknikleri kullanılarak cilt lezyonları için çok sınıflı sınıflandırma

Abstract

Dünya çapında en yaygın kanser türlerinden biri olan cilt kanseri, erken teşhis edilmediğinde ölümcül sonuçlara yol açabilir. Bu durumda, cilt lezyonlarının erken evrede doğru bir şekilde tespit edilmesi ve sınıflandırılması, etkili bir tedavi süreci için çok önemlidir. Dermatologlar, geleneksel tanı süreçlerini manuel olarak yürütürler, bu da zaman kaybına, maliyetlere ve subjektif farklılıklara neden olabilir. Yapay zeka teknolojisindeki ve özellikle derin öğrenme teknolojilerindeki gelişmeler, deri lezyonlarını yüksek doğrulukla ve otomatik bir şekilde sınıflandırmayı mümkün kılmıştır. Bu tez çalışmasının amacı, Residual Neural Network mimarisine sahip ResNet-18 modelinin kullanılarak deri lezyonlarının çok sınıflı sınıflandırmasının yapılmasıdır. Çalışmada, uluslararası alanda yaygın olarak kullanılan ve açık erişime sunulan ISIC 2019 veri seti tercih edilmiştir. Bu veri seti, sekiz farklı cilt lezyonu türünü içermekte olup; aktinik keratoz, bazal hücreli karsinom, iyi huylu keratoz, dermatofibroma, melanositik nevüs, skuamöz hücreli karsinom, melanom ve vasküler lezyon gibi yaygın deri hastalıklarına ait 25.331 etiketlenmiş görüntüyü kapsamaktadır. Ancak, veri setinde yer alan sınıflar arasında belirgin bir veri dengesizliği bulunması nedeniyle bu çalışmada üç ana cilt lezyonu türü (bazal hücreli karsinom, iyi huylu keratoz ve melanom) üzerinde sınıflandırma yapılmıştır. Bu seçim, sınıflar arasındaki örnek sayılarının birbirine daha yakın olmasına ve modelin eğitim sürecinde daha dengeli bir öğrenme gerçekleştirebilmesine olanak sağlamıştır. Modelin başarısını artırmak için görüntü boyutlarını değiştirdik, çeşitli veri artırma tekniklerini kullandık ve eğitim veri setlerini dengeledik. Veriyi artırmak için yeniden boyutlandırma, yatay ve dikey çevirme, döndürme ve renk kontrastı gibi çeşitli teknikler kullanılmıştır. Veri kümesinin %80'i eğitim seti, %10'u doğrulama seti ve %10'u test seti olarak ayrılmıştır. Transfer öğrenimi, modeli ResNet'ten önceden eğitilmiş ağırlıklarla başlatılarak ve ardından cilt lezyonu veri kümesinde ince ayar yapılarak kullanılmıştır. Model, AdamW optimize edicisi kullanılarak optimize edilmiştir. Model, 64'lük bir mini küme (mini batch) boyutu ve 0,0001'lik bir başlangıç öğrenme oranı ile 100 dönemde eğitildi. Çapraz entropi kaybı, kayıp fonksiyonu olarak kullanılır. Kayıp fonksiyonu, eğitim boyunca tahmin edilen değerler ile gerçek değerler arasındaki farkı niceliksel olarak belirler. Daha düşük kayıp değerleri, modelin daha iyi çalıştığını gösterir. Model eğitimi ve test aşamalarında çapraz doğrulama tekniği kullanılarak, sınıflandırma performansı, F1 skorları, doğruluk (doğruluk), hassasiyet (hassasiyet), duyarlılık (duyarlılık), özgüllük (özgüllük) ve diğer kriterlere göre değerlendirilmiştir. Çalışmanın sonucunda, cilt lezyonlarını sınıflandırma modelimizi kullanarak test setinde %82,26 doğruluk, %82,31 kesinlik, %82,26 duyarlılık ve %82,27 F1 skorları elde ettik. Bu çalışma ResNet-18 modelinin düşük parametre sayısı sayesinde daha hızlı eğitilebilir olduğunu ve çok sınıflı cilt lezyonu sınıflandırmasında başarılı bir performans sergilediğini göstermiştir. Sonuç olarak bu çalışma, derin öğrenme tabanlı modellerin tıbbi görüntü analizi alanında başarılı bir şekilde uygulanabileceğini ve özellikle cilt lezyonlarının bilgisayar destekli teşhis etmek için büyük bir potansiyele sahip olduğunu göstermektedir. Geliştirilen sistemin, ilerleyen süreçlerde klinik karar destek sistemlerine entegre edilerek dermatologlara yardımcı olması ve erken teşhis süreçlerine katkı sunması hedeflenmektedir.

One of the most prevalent cancers in the world is skin cancer, can lead to fatal outcomes if not diagnosed early. In this case, accurate detection and classification of skin lesions at an early stage is crucial for effective treatment. Dermatologists conduct traditional diagnostic processes manually, which can result in loss of time, costs, and subjective differences. Developments in artificial intelligence technology and especially deep learning technologies have made it possible to classify skin lesions with high accuracy and automatically. The purpose of this thesis was to perform multi-class classification of skin lesions using the ResNet-18 model with Residual Neural Network architecture. The ISIC 2019 dataset, which is widely used internationally and openly accessible, was preferred in the study. This dataset includes eight different skin lesion types; it includes 25.331 labeled images of common skin diseases such as actinic keratosis, basal cell carcinoma, benign keratosis, dermatofibroma, melanocytic nevus, squamous cell carcinoma, melanoma and vascular lesions. However, due to a significant data imbalance between the classes in the dataset, classification was made on three main types of skin lesions (basal cell carcinoma, benign keratosis and melanoma) in this study. This selection enabled the number of examples in each class to be more equal, allowing the model to achieve a more balanced learning experience during training. To improve performance of the model, we changed the image sizes, applied different data augmentation methods, and balanced the training datasets. Various techniques such as resizing, horizontal and vertical flipping, rotation, and color contrast were used to augment the data. Images were resized to ensure consistent input dimensions for the Res-Net. Both horizontal and vertical flips were applied to generate new training samples, enhancing the model's ability to recognize lesions from different orientations. Random rotations were performed on the images to simulate various real-world viewing angles. Changes in brightness, contrast, and saturation were introduced to make the model invariant to lighting conditions. These data augmentation methods help to artificially expand the training set, ensuring that the model learns to generalize better from diverse variations of skin lesion images. The dataset was divided into three different subsets for training, validation, and testing the model. This division is a critical step to accurately assess the generalization capabilities of machine learning and deep learning models. In this study, 80% of the dataset was divided into training data, 10% as validation data, and the remaining 10% as test data. While the training data was used throughout the learning process of the model, the validation data was used for evaluation purposes to monitor the performance of the model during the training process and to prevent overfitting. The test data was used after the model was trained to evaluate its overall performance on previously unseen data. Transfer learning approach was adopted in training the model. In this method, the weights of a model previously trained on a large and comprehensive dataset are taken as the starting point. In this study, pre-trained weights of ResNet (Residual Network) model, one of the widely used deep convolutional neural network architectures, were used. Thanks to these pre-trained weights, the model has already learned low-level visual features (edges, colors, textures, etc.), which allows it to learn faster and more effectively on the target dataset. Then, the model is initialized with these weights and continues to be trained by fine-tuning on the skin lesions dataset. An advanced optimization algorithm called AdamW (Adaptive Moment Estimation with Weight Decay) was used for the optimization of the model. AdamW is a variant of the classical Adam optimization algorithm and is known for performing the weight decay process more accurately. In this way, the model generalizes better. AdamW provides both rapid convergence with momentum-like terms and adaptive learning rates and prevents over-learning by preventing excessive growth of parameters. In this study, the initial learning rate was determined as 0.0001. This low learning rate allows the model to learn more steadily and reach the optimal weights more stably. The training of the model was performed in mini-batches consisting of 64 samples. The mini-batch method allows the model to be updated in small groups instead of processing the entire dataset at once. This is both advantageous in terms of memory usage and helps the model learn faster. During the training process, the model was run on the entire dataset for 100 epochs. An epoch is when the model processes the entire training data once. That is, in each epoch, the model passes over the entire training data once. Repeating this process 100 times provides ample opportunity for the model to learn patterns in the data and reach higher levels of accuracy. The cross-entropy loss function was used to evaluate the accuracy of the model's predictions. This function is one of the most commonly used loss functions in classification problems. Cross-entropy loss calculates the difference between the model's predicted class probability distribution and the actual class labels. This difference between the actual and predicted values is expressed numerically as a "loss" value. The smaller this value, the more accurate the model's predictions are. During the training process, the model's weights are updated to minimize this loss value. Thus, the model starts to make more accurate predictions over time. Using the cross-validation technique in the model training and testing stages, the classification performance was evaluated according to F1 scores, accuracy, precision, sensitivity, specificity and other criteria. As a result of the study, we achieved 82.26% accuracy, 82.31% precision, 82.26% recall and 82.27% F1 scores on the test set using our skin lesion classification model. This study showed that the ResNet-18 model can be trained faster thanks to the low number of parameters and has a successful performance in multi-class skin lesion classification. This study demonstrates the significant promise of deep learning techniques, particularly residual neural networks, in supporting dermatological diagnoses through automated classification of skin lesions. Such AI-driven solutions have the potential to alleviate the workload on dermatologists, shorten the time needed for diagnosis, and increase accuracy, especially in environments with limited medical resources. Although this research concentrated on three specific lesion categories, broadening the scope to cover additional lesion types and incorporating relevant clinical data could further improve model effectiveness and its applicability in clinical practice. Future research directions may involve investigating more sophisticated deep learning frameworks, leveraging ensemble learning strategies, and combining image analysis with patient information to create multimodal diagnostic tools. Furthermore, integrating these models into clinical decision support platforms and conducting validation studies in actual healthcare settings with dermatologist involvement would be critical steps towards real-world implementation. In summary, this thesis confirms that the ResNet-18 architecture, enhanced by transfer learning and comprehensive data augmentation, can accurately classify multiple types of skin lesions based on the ISIC 2019 dataset. These results contribute to the advancement of AI-based diagnostic tools aimed at facilitating early skin cancer detection, ultimately supporting improved patient outcomes and healthcare delivery.