In this paper, we discuss and try a multi-image fusion approach with a multibranch Convolutional Neural Network (CNN) that implies a MultiHeadAttention (MHA) technique in the fusion center. This work studies the employment of the architecture on actual data containing different images of the same USB device. The images differ in the direction of the light at the moment of the acquisition. We observed that instead of employing a simple concatenation fusion of the outputs, the network architecture could employ a multibranch classification featurewise, which utilizes a multi-head attention mechanism instead of a channel attention one.

In recent years, segmentation algorithms utilizing deep learning have achieved outstanding performance in medical image segmentation. However, accurately delineating small anatomical structures continues to be a challenging task, even for the most advanced methods that produce impressive results. This challenge might arise from the use of small kernels and downsampling operations, which often emphasize complex high-level features at the expense of low-level details like edges. Inspired by recent research highlighting this challenge, we developed a novel architecture that combines the standard U Net with an additional branch harnessing the potential of large convolutional kernels. These large kernels are utilized in a decreasing-increasing manner over image features of the same size, guiding the network to focus on smaller parts. The proposed method demonstrated strong potential in segmenting small anatomical structures, surpassing our baseline and matching the performance of a robust state-of-the-art network across various datasets and domains, all while maintaining a relatively small number of parameters.

The majority of evidential fusion models presented in the literature is based on optimistic assumptions about the reliability of the models producing beliefs and assumes that they are equally reliable. At the same time, the belief models used in combination may have some limitations and may result in different reliabilities, which may decrease the performance of the combination. One way to confront this problem is to consider a discount rule utilizing reliability coefficients. One of the problems of using discounting is the way of modeling reliability coefficients. This paper proposes modeling reliability coefficients by considering a new effective measure of belief uncertainty. The new reliability coefficients are introduced in a multilayer decision fusion-based Convolutional Neural Network (CNN) architecture built within the Transferable Belief Model, as well as in a multimodal deep learning scenario. Case study results demonstrate the feasibility of representing reliability by the belief uncertainty measure considered.