Speaker
Description
Thyroid cancer is a prevalent form of endocrine cancer that requires early diagnosis to optimize patient outcomes. Importantly, standard diagnostic procedures, such as fine-needle aspiration biopsy, may be invasive. For this reason, it is crucial to explore new technologies to detect this pathology. In this context, infrared thermography (IRT) is a non-invasive imaging method that may evaluate thermal abnormalities linked to diseased processes, such as neoplastic forms. In fact, the metabolic hyperactivity and abnormal vascularization characteristic of malignant thyroid nodules generate peculiar thermal patterns on the skin surface, which may be identified using deep learning (DL) algorithms. Moreover, using transfer learning methods, which use pre-trained models on large and general datasets, allows handling small datasets while still getting good classification accuracy.
The objective of this work was to create a two-class classifier for thyroid tumors (carcinoma vs adenoma) using IRT images by transfer learning with the VGG-16 architecture. In this study, a publicly accessible dataset (https://doi.org/10.3390/s21134459), consisting of thermal pictures from 97 patients (61 with adenoma and 36 with carcinoma), aged 25-75 years, acquired with a FLIR SC620 thermal camera (640×480 pixel resolution). The images were resized to 224×224 pixels and normalized into the [0,1] range. The dataset was split into 70% for training and validation (5-fold) and 30% for testing. Notably, the most numerous class was randomly undersampled to guarantee the balance between the two classes. The VGG-16 model was built with frozen convolutional layers to keep learned generic features. Classification layers were added, such as a global average pooling layer, a fully connected layer with 128 neurons and ReLU activation, a dropout layer (rate=0.5) for regularization, and a sigmoid output layer for binary classification. During training, data augmentation techniques were employed including rotation (±20°), width and height change (10%), shearing (10%), zoom change (20%), brightness change (0.9–1.1 range), and horizontal flipping of image. The model was compiled using the Adam optimizer and binary cross-entropy loss. This model was trained for up to 50 epochs with a batch size of 20. In order to avoid overfitting, an early stopping with a patience of 10 epochs and validation loss was employed. The model showed an overall accuracy of 93.2%, precision of 94.1% and recall of 92.3%. The results suggest that transfer learning with VGG-16 system might be able to efficiently identify the thermal patterns related to the thyroid tumors. The non-invasive nature of IRT combined with DL methods offers clinical potential as a large-scale screening tool, particularly useful in resource-limited areas or for patients requiring frequent monitoring. Further research including multi-centric investigations, using larger more diverse datasets and multimodal imaging approaches is indeed needed to enhance the diagnostic capability of the method.
