Refine
Document Type
Language
- English (2)
Has Fulltext
- no (2)
Is part of the Bibliography
- no (2)
Keywords
- Deep learning (1)
- Lightweight Architectures (1)
- Skin Lesion Detection (1)
- Trans- formers (1)
Institute
Begutachtungsstatus
- peer-reviewed (2)
LiwTERM: A Lightweight Transformer-Based Model for Dermatological Multimodal Lesion Detection
(2024)
Skin cancer is the most common type of cancer in the world, accounting for approximately 30% of all diagnosed tumors. Early diagnosis reduces mortality rates and prevents disfiguring effects in different body regions. In recent years, machine learning techniques, particularly deep learning, have shown promising results in this task, presenting studies that have demonstrated that combining a patient’s clinical information with images of the lesion is crucial for improving the classification of skin lesions. Despite that, meaningful use of clinical information with multiple images is mandatory, requiring further investigation. Thus, this project aims to contribute to developing multimodal machine learning-based models to cope with the skin lesion classification task employing a lightweight transformer model. As a main hypothesis, models can take multiple images from different sources as input, along with clinical information from the patient’s history, leading to a more reliable diagnosis. Our model deals with the not-trivial task of combining images and clinical information (from anamneses) concerning the skin lesions in a lightweight transformer architecture that does not demand high computation resources but still presents competitive classification results.
Barrett’s esophagus, also known as BE, is commonly associated with repeated exposure to stomach acid. If not treated properly, it may evolve into esophageal adenocarcinoma, aka esophageal cancer. This paper proposes TransConv, a hybrid architecture that benefits from features learned by pre-trained vision transformers (ViTs) and convolutional neural networks (CNNs), followed by a shallow neural network composed of three normalizations, ReLU activations, and fully connected layers, and a SoftMax head to distinguish between BE and esophageal cancer. TransConv is designed to be training-lightweight, and for the ViT and CNN backbone models, weights are kept frozen during training, i.e., the primary goal of TransConv is to learn the weights of the fully connected layer from both backbones only, avoiding the burden of updating their weights but still learning their final descriptions for the lightweight convolutional model. We report promising results with low computational training costs in two datasets, one public and another private. From our achievements, TransConv was able to deliver balanced accuracy results around 85% and 86% for each evaluated dataset, respectively, in a design that required only 50 epochs of model training, a very reduced number compared to state-of-the-art conducted studies in the same domain.