Projects

This project aims to automate the classification of immunological phenotypes in human solid tumors, crucial for predicting disease progression and tailoring immunotherapy. We developed an end-to-end pipeline for histology image analysis to identify and classify stromal, lymphocyte, and cancer cells. The workflow includes cell segmentation using HoVer-Net and classification using an AutoML-based ensemble, achieving a weighted F1-score of 79.8%. The pipeline also categorizes tumors into immune phenotypes (desert, excluded, inflamed, or unknown) with high accuracy. The proposed solution provides a robust, efficient tool for medical laboratories, enabling quick and reliable analysis of tissue samples to support cancer diagnosis and treatment.

This project aims to advance histological analysis of tissue-engineered vascular grafts (TEVGs) using machine learning (ML) tools. We compiled a dataset of 104 whole slide images (WSIs) from TEVGs implanted in sheep carotid arteries for six months and annotated 1401 patches to identify nine key histological features. Six deep learning models, including U-Net and MA-Net, were evaluated for their segmentation performance. MA-Net achieved the highest mean Dice Similarity Coefficient (DSC) of 0.875, and an ensemble of top-performing models reached an average DSC of 0.889. This ML-driven approach enhances the accuracy and efficiency of histopathological analysis, providing a valuable tool for tissue engineering research.

This project pioneers a novel method for designing prosthetic heart valves (PHVs) by integrating machine learning and optimization algorithms. By combining parametric modeling and Finite Element Method simulations, this approach leverages a comprehensive dataset of 11,565 PHV designs to predict key performance metrics like lumen opening area and peak stress. The best-performing machine learning models, enhanced by advanced optimization algorithms, achieved high accuracy and computational efficiency. Notably, the Tree-structured Parzen Estimator and Nondominated Sorting Genetic Algorithm effectively balanced design parameters, marking a significant advancement in creating more effective and durable heart valves.

This study presents a deep learning method for diagnosing pulmonary edema from chest radiographs, crucial for managing congestive heart failure. Using 1000 annotated radiographs, the method involves lung segmentation followed by feature localization. The Side-Aware Boundary Localization network excelled, achieving a mean average precision of 0.568 for detecting key edema features. This approach promises an accurate, efficient, and interpretable tool for assessing pulmonary edema severity, enhancing diagnostic precision in clinical care.

This study describes a new two-stage workflow for the segmentation and scoring of lung diseases on X-ray images. The workflow includes two core stages for lung and disease segmentation, as well as a post-processing stage for scoring. The models used in this workflow were trained and tested on four publicly available X-ray datasets of COVID-19 patients and two X-ray datasets of patients with no pulmonary pathology. The best-performing algorithm, which is based on a combination of DeepLabV3+ for lung segmentation and MA-Net for disease segmentation, had a mean absolute error of 0.3 out of 6.0, which was significantly lower than established COVID-19 algorithms. Additionally, this workflow was also more computationally efficient, taking about 11 times less time than the other methods. This approach has the potential to be used for other lung diseases such as pneumonia, tuberculosis, and pneumothorax.

In this study, we have developed a model that can search for and anonymize faces and ears in CT and MRI data, leaving the rest of the brain untouched. The model uses image processing algorithms such as mosaicing or blurring to protect patient privacy and was trained and tested on a dataset of 551 CT series and 555 MRI series. The model achieved an average precision of 74.1% on CT data and 92.6% on MRI data, with a processing time of 37 ms per image on an NVIDIA RTX 3090 GPU and 880 ms per image on an AMD Ryzen Threadripper 3960X GPU.

This study aims to stop the spread of cancer at its earliest stages by estimating the spread of different types of cancer cells. We developed a pipeline for counting and localizing primary nuclei, micronuclei, mitosis, and apoptosis in cancer cells. We used the EfficientDet network pre-trained on the ImageNet dataset. This architecture includes a BiFPN feature network that takes level 3-7 features from the backbone network and applies top-down and bottom-up bidirectional feature fusion. The dataset used for this study was collected from the internal Volastra dataset, the Image Data Resource, and the HMS LINCS database, and included 36 cell lines representing 1664 images. The results showed that the network had a mean average precision (mAP) of 66% for micronuclei and 85% mAP for primary nuclei.

The study describes the use of modern machine learning methods, specifically a CNN-based method, for the detection of COVID-19 using patients' chest X-rays. The proposed method utilizes indirect supervision based on Grad-CAM, which uses attention heatmaps to support the network's predictions during the training process. We combine publicly available data from five different sources and annotate the images for normal, pneumonia, and COVID-19 to increase classification accuracy. We also propose a training pipeline based on indirect supervision of traditional classification networks, where guidance is provided by an external algorithm. As a result, we found that standard networks can achieve an accuracy comparable to custom models for COVID-19, with one network in particular, VGG-16, outperforming the best of the tailor-made models.

The study discusses a proposed multi-task learning algorithm for tracking the location of anatomical landmarks and labeling critical keypoints on both the aortic valve and delivery system during transcatheter aortic valve implantation (TAVI) procedures. The algorithm uses 9 neural networks based on various architectures to predict 11 keypoints of interest. The ResNet V2 and MobileNet V2 architectures were found to have the best prediction accuracy and speed, with 97 and 96% accuracy and 4.7 and 5.6% mean absolute error, respectively. The algorithm is intended to improve the accuracy of valve positioning during TAVI procedures.

This study aims to confirm the feasibility of using deep learning methods for real-time detection of coronary artery stenoses in invasive coronary angiography, which is the gold standard for diagnosing coronary artery disease. Eight promising detectors based on different neural network architectures were trained and tested on clinical angiography data from 100 patients. The results showed that three neural networks performed best, with the most accurate network achieving a mean average precision of 0.95 and an F1 score of 0.96 with a prediction rate of 3 fps. Another network was the fastest, with a prediction rate of 38 fps, but had lower accuracy. The study confirms that deep learning methods can be used for real-time detection of coronary artery stenosis, which can aid in the decision-making process when interpreting angiography findings.

The study describes a project that uses deep learning to identify and locate wildfires in the Siberian wild forest. The system was developed in response to active wildfires in 2019 and is intended to alert the Forest Protection Service. The system is based on an object detection network that uses open-source datasets from various universities and organizations to train the network. The system was tested on different object detectors and found that the EfficientDet-D1 detector was the most efficient in terms of localization accuracy, achieving a classification accuracy of 81% and a localization accuracy of 87%. The processing speed of the network was found to be sufficient, processing images at a rate of approximately 9 frames per second.

This study deals with the segmentation and visualization of catheters during minimally invasive surgery. A modification of the U-net architecture, called V-net, was used to solve the problem of segmenting objects on images with speckle noise. The V-net has additional skip-connections inside the encoder and decoder, which improves the segmentation accuracy, specifically the Dice Similarity Coefficient. The V-Net was found to be 10% more accurate than a classical U-Net. The dataset used in this study was obtained by epicardial three-dimensional echocardiography during cardiac surgery on three Yorkshire pig hearts at Boston Children's Hospital. A total of 75 three-dimensional ultrasound grayscale samples were acquired using the Philips iE33 ultrasound machine and PMS5.1 ultrasound software. The catheter is poorly visible on the echocardiographic data, so it is highlighted with green circles and ellipses. This study is related to the project "Segmentation of anatomical structures".

In this study, a ray-casting-based segmentation algorithm for boundary detection was developed. The resolution of the object boundary is determined by the number of rays cast, which varies depending on the shape of the target region. The algorithm was tested on a cardiac MRI dataset from the University of York and a brain tumor dataset from Southern Medical University. The highest Dice similarity coefficients for heart and brain tumor segmentation were 86.5% and 89.5%, respectively. The processing time for a heart image was 4.1 ms for 8 slices and 20.2 ms for 64 slices. The algorithm is fast, highly accurate for convex and closed objects, and can be scaled to incorporate different boundary detection techniques.

The article describes a proposed machine learning solution for the segmentation of various anatomical structures, such as the left atrium, pancreas, spleen, hippocampus, and liver. The solution is based on a neural network operating in 3D mode, using a V-net architecture with additional skip connections to partially solve the problem of vanishing gradients. The proposed V-net was compared to 11 existing approaches in the "Medical Segmentation Decathlon" competition and achieved the best segmentation accuracy on 3 out of 5 tested structures, and the second and seventh best accuracy on the other two structures. The project is related to "Segmentation of medical devices for minimally invasive surgery".