AB025. Machine learning based image classification in neutron autoradiography

Background: Neutron autoradiography can be used to address boron concentration in different tissular structures and the homogeneity of tumour dose-distribution. By irradiating a boronated sample with a thermal neutron flux, the reaction ¹⁰B(n,α)⁷Li (BNC) takes place, leaving a localized damage zone in the nuclear track detector (NTD). After a chemical process, tracks can be observed under optical microscopy and translated into a boron concentration value or distribution. However, accurate results depend on the correct track density quantification, which requires controlled image acquisition and processing conditions. To address this issue, an image classification algorithm was developed using feature extraction to reject inadequate images.

Methods: The training set consisted of BNC images acquired over several years, corresponding to aqueous solutions and biological tissues with different ¹⁰B concentrations. A label “Accepted” or “Rejected” was defined for each image, and a set of morphological and uniformity parameters were extracted from the quantified objects. Statistical parameters were computed for each characteristic, so 36 features, plus a label, were used to represent each image. A series of machine learning models were evaluated for this application, and those with the highest performance were a support vector machine (SVM) and an artificial neural network (NN). The SVM algorithm was trained with a radial basis function (RBF) kernel and using cross-validated randomized search for finding the best hyperparameters. The NN was trained several times adjusting the number of neurons, layers, optimization algorithms, learning rate, among others.

Results: The final performance metrics turned out to be comparable for both models: 93% for both accuracy and precision for the SVM, vs. 94% accuracy and 95% precision for the NN, with the NN showing better feature learning capacity based on the distribution of predicted class probabilities. Additionally, these results did not show significant variations when analysed according to track density, thus, the NN was selected to perform the image verification step prior to quantification.

Conclusions: This approach shows the potential of including machine learning methods in the autoradiographic analysis, so we plan to extend their use to other steps in the workflow.

Keywords: Nuclear tracks; machine learning; neutron autoradiography