MIST: Multi-instance selective transformer for histopathological subtype prediction.

Accurate histopathological subtype prediction is clinically significant for cancer diagnosis and tumor microenvironment analysis. However, achieving accurate histopathological subtype prediction is a challenging task due to (1) instance-level discrimination of histopathological images, (2) low inter-class and large intra-class variances among histopathological images in their shape and chromatin texture, and (3) heterogeneous feature distribution over different images. In this paper, we formulate subtype prediction as fine-grained representation learning and propose a novel multi-instance selective transformer (MIST) framework, effectively achieving accurate histopathological subtype prediction. The proposed MIST designs an effective selective self-attention mechanism with multi-instance learning (MIL) and vision transformer (ViT) to adaptive identify informative instances for fine-grained representation. Innovatively, the MIST entrusts each instance with different contributions to the bag representation based on its interactions with instances and bags. Specifically, a SiT module with selective multi-head self-attention (S-MSA) is well-designed to identify the representative instances by modeling the instance-to-instance interactions. On the contrary, a MIFD module with the information bottleneck is proposed to learn the discriminative fine-grained representation for histopathological images by modeling instance-to-bag interactions with the selected instances. Substantial experiments on five clinical benchmarks demonstrate that the MIST achieves accurate histopathological subtype prediction and obtains state-of-the-art performance with an accuracy of 0.936. The MIST shows great potential to handle fine-grained medical image analysis, such as histopathological subtype prediction in clinical applications.

Improvement of Phased Antenna Array Applied in Focused Microwave Breast Hyperthermia.

Focused microwave breast hyperthermia (FMBH) employs a phased antenna array to perform beamforming that can focus microwave energy at targeted breast tumors. Selective heating of the tumor endows the hyperthermia treatment with high accuracy and low side effects. The effect of FMBH is highly dependent on the applied phased antenna array. This work investigates the effect of polarizations of antenna elements on the microwave-focusing results by simulations. We explore two kinds of antenna arrays with the same number of elements using different digital realistic human breast phantoms. The first array has all the elements' polarization in the vertical plane of the breast, while the second array has half of the elements' polarization in the vertical plane and the other half in the transverse plane, i.e., cross polarization. In total, 96 sets of different simulations are performed, and the results show that the second array leads to a better focusing effect in dense breasts than the first array. This work is very meaningful for the potential improvement of the antenna array for FMBH, which is of great significance for the future clinical applications of FMBH. The antenna array with cross polarization can also be applied in microwave imaging and sensing for biomedical applications.

Deep-Learning-Enabled Microwave-Induced Thermoacoustic Tomography Based on ResAttU-Net for Transcranial Brain Hemorrhage Detection.

OBJECTIVE: Hemorrhagic stroke is a leading threat to human's health. The fast-developing microwave-induced thermoacoustic tomography (MITAT) technique holds potential to do brain imaging. However, transcranial brain imaging based on MITAT is still challenging due to the involved huge heterogeneity in speed of sound and acoustic attenuation of human skull. This work aims to address the adverse effect of the acoustic heterogeneity using a deep-learning-based MITAT (DL-MITAT) approach for transcranial brain hemorrhage detection. METHODS: We establish a new network structure, a residual attention U-Net (ResAttU-Net), for the proposed DL-MITAT technique, which exhibits improved performance as compared to some traditionally used networks. We use simulation method to build training sets and take images obtained by traditional imaging algorithms as the input of the network. RESULTS: We present ex-vivo transcranial brain hemorrhage detection as a proof-of-concept validation. By using an 8.1-mm thick bovine skull and porcine brain tissues to perform ex-vivo experiments, we demonstrate that the trained ResAttU-Net is capable of efficiently eliminating image artifacts and accurately restoring the hemorrhage spot. It is proved that the DL-MITAT method can reliably suppress false positive rate and detect a hemorrhage spot as small as 3 mm. We also study effects of several factors of the DL-MITAT technique to further reveal its robustness and limitations. CONCLUSION: The proposed ResAttU-Net-based DL-MITAT method is promising for mitigating the acoustic inhomogeneity issue and performing transcranial brain hemorrhage detection. SIGNIFICANCE: This work provides a novel ResAttU-Net-based DL-MITAT paradigm and paves a compelling route for transcranial brain hemorrhage detection as well as other transcranial brain imaging applications.

Deep Transfer Learning-Based Multi-Object Detection for Plant Stomata Phenotypic Traits Intelligent Recognition.

Plant stomata phenotypic traits can provide a basis for enhancing crop tolerance in adversity. Manually counting the number of stomata and measuring the height and width of stomata obviously cannot satisfy the high-throughput data. How to detect and recognize plant stomata quickly and accurately is the prerequisite and key for studying the physiological characteristics of stomata. In this research, we consider stomata recognition as a multi-object detection problem, and propose an end-to-end framework for intelligent detection and recognition of plant stomata based on feature weights transfer learning and YOLOv4 network. It is easy to operate and greatly facilitates the analysis of stomata phenotypic traits in high-throughput plant epidermal cell images. For different cultivars, multi-scales, rich background features, high density, and small stomata object images, the proposed method can precisely locate multiple stomata in microscope images and automatically give phenotypic traits of stomata. Users can also adjust the corresponding parameters to maximize the accuracy and scalability of automatic stomata detection and recognition. Experimental results on actual data provided by the National Maize Improvement Center show that the proposed method is superior to the existing methods in high stomata automatic detection and recognition accuracy, low training cost, strong generalization ability.