Deep Learning for Lymphoma Detection on Microscopic Images
Ammar Ammar, Irfan Tito Kurniawan, Resfyanti Nur Azizah, Hafizh Rahmatdianto Yusuf, Antonius Eko Nugroho, Ghani Faliq Mufiddin, Isa Anshori, Widyawardana Adiprawita, Hermin Aminah Usman, Okky Husain · 2022
Early lymphoma diagnosis is essential to improve the patients' survival rate and avoid irreversible damage.Immunohistochemistry-based lymphoma diagnostics is an expensive and time-consuming process, especially in developing countries with limited resources.Image-based lymphoma diagnostics might serve as an inexpensive, yet less accurate alternative to immunohistochemistry-based methods.One challenge in image-based methods is that carcinoma can occur in the same organ as lymphoma, thus making it hard to differentiate the two types of cancer.To assist lymphoma diagnostics, this study proposes a deep learningbased method to classify nasopharyngeal microscopic biopsy images into one of three classes: lymphoma, carcinoma, and benign lesion.The method works by splitting the images into patches, classifying each patch using a deep learning model, and taking the average confidence score of each patch.We compared three deep learning-based feature extractor architectures and studied the effects of three image color preprocessing techniques on classification performance.We reached 88.7% sensitivity and 91.3% specificity in differentiating lymphoma on 400x magnification CLAHE-enhanced microscopic images using the InceptionResNetV2 model.We also reached 87.0% three-class classification accuracy using the same model.