Bulat Maksudov

Bulat Maksudov

29 y.o. PhD Researcher at DCU. Passionate about Explainable Medical AI.

Posts by Collection

portfolio

publications

Increase of efficiency of relay protection reliability in modes with deep saturation of current transformers using the methodology based on the application of artificial neural networks

Published in 2018, 2018

There are different types of tasks in power engineering, which are solved by means of hardware and software complexes [1]. Some of the issues might be resolved within short period using relatively simple algorithms. Other issues require collecting data for estimating tendency of changes in condition of operating equipment by mathematical models. There are tasks, which cannot be solved with the help of mathematical model’s synthesis, which is based on the passport or other data about the equipment [2]. This kind of tasks consists of forecasting the technical condition of the equipment for maintenance planning according to the actual equipment state, forecasting of emergency situations, classification of operating modes of the electric power facilities, different types of restoration of initial signals etc.

Software Package For Improving Financial And Technological Performance Of Microgrid Networks

Published in 2019, 2019

In modern world, the use of the most advanced digital technologies in any industry and business directly effect on financial and technological indicators, in other words, the more advanced digital technologies are used to solve various tasks, the greater the profit or benefits that can be gained. The electric power industry is not an exception. Nowadays, more and more of the electric power sector is moving from large networks to small, often isolated, so-called Microgrid. Such networks generally have generation based on renewable energy sources (RE): wind power plants, solar power plants, small hydropower plants, tidal power plants, etc.In view of the fact that generation is stochastic, networks with generation relying on renewable energy sources have energy storage. It is also worth noting that isolated and non-isolated Microgrid even if generation does not base on renewable energy, energy storage devices can bring them a certain benefits connected with the changing cost of electricity during the day or year. Thus, a consumer of Microgrid has much greater capabilities than a consumer of a “traditional”, centralized power supply system, but he is also subject to much greater responsibility, because many of the functions that the system operator used to perform now fall on his shoulders. So, the following customer features in Microgrid can be distinguished: the ability to disconnect themselves from the mains supply for the period when consuming is not profitable for them, the ability to sell electricity to the power supply network, independently maintain equipment (including generating), calculate and forecast their consumption and generation, make profit from the sale of electricity to the network.Obviously, in the past, the average consumer was not capable meeting the greatest part of the needs of his own electrical “industry” independently, but using modem digital technologies, most of the tasks that previously were impossible to fulfill could be automated without the direct participation of the consumer. That is why, it is proposed to use software systems which will be based on neural networks. The task of these software systems is to collect and process monitoring data, each consuming or generating unit in Microgrid, to perform a large number of tasks. One of such tasks is classification and creation characteristics of generating and consuming equipment by collecting and analyzing data from Microgrid participants. The algorithm determines the characteristics of consumption and generation. Based on these characteristics under various external conditions and factors, load and generations schedules (especially important for renewable energy sources) and possible emergency events are predicted. In addition, such software systems make it possible to optimize the algorithms for determining the most profitable hours for consumption or selling of electricity to the network. It is more convenient to operate the described software systems as a cloud services. In other words, in order to start implement a software package into operation, the consumer will only need an Internet connection, so there is no need for computers with high computational abilities, all calculations occur remotely.This work describe a software package which include the automation, forecasting and optimization of the financial and technological performance of Microgrid networks. It considers the data that the software package needs for complete analysis and further prediction, methods and algorithms that underpin this software package and the possible benefit from its use. RTDS hardware and software system was used to model the power system; the prediction methodology was based on recurrent neural networks (RNN).Machine learningWhat exactly is machine learning? It is obvious that “learning” is when a certain model” learns”in a some way and then begins to return results, that is, most likely, to predict something. A very general definition of”learnability” is roughly the same as that given by Thomas Mitchell in his book “Machine learning” [4]: “A computer program is said to learn from experience with some class of tasks T and performance measure P, if its performance at tasks (as measured by P) improves with experience” [5].The main classification of machine learning tasks is shown in Fig. 1. The two main classes of machine learning tasks are supervised learning and unsupervised learning. To fulfill the purposes-to detect faults in power transformers on the basis of PMU, it is necessary to carry out supervised learning tasks such as data classification. In the work presented here, data is a set of features that will be fed to the neural network input with the expected output: “true”- interwinding fault occurrence, “false”- without inter-winding fault. In order for a trained neural network to accurately detect turn-to-turn faults in a transformer it is highly important to prepare a sufficient set of data and to select the features of this type of damage as accurately as possible.

Task-Oriented Language Grounding for Language Input with Multiple Sub-Goals of Non-Linear Order

Published in arXiv preprint, 2019

In this work, we analyze the performance of general deep reinforcement learning algorithms for a task-oriented language grounding problem, where language input contains multiple sub-goals and their order of execution is non-linear. We generate a simple instructional language for the GridWorld environment, that is built around three language elements (order connectors) defining the order of execution: one linear - “comma” and two non-linear - “but first”, “but before”. We apply one of the deep reinforcement learning baselines - Double DQN with frame stacking and ablate several extensions such as Prioritized Experience Replay and Gated-Attention architecture. Our results show that the introduction of non-linear order connectors improves the success rate on instructions with a higher number of sub-goals in 2-3 times, but it still does not exceed 20%. Also, we observe that the usage of Gated-Attention provides no competitive advantage against concatenation in this setting. Source code and experiments’ results are available at https://github.com/vkurenkov/language-grounding-multigoal

Adopting Confident Learning to Eliminate Uncertainty in Chest X-ray Images for Lung Nodules Prediction

Published in 2020 ASTRO Annual Meeting, 2020

The potential of deep learning to advance lung nodule detection from chest X-rays is significantly compromised by the lack of large annotated databases and noisy labels in the existing databases. The aim of this study is to investigate the applicability of the novel Confident Learning approach for chest X-ray database cleaning and nodule detection improving.

Download Paper

Automated Localization of Lung Nodules from Chest X-rays With Deep Neural Networks

Published in 2020 ASTRO Annual Meeting, 2020

Lung cancer constitutes more than 20% of all cancer deaths in the Russian Federation. About 34.2% of these cases were diagnosed late, which significantly reduces the life expectancy of patients. Chest X-rays are the main screening method for lung cancer in Russia. The small size of nodules and difficult localizations are the reasons why nodules are often missed during routine scanning. The aim of this study is to employ modern deep learning tools for the localization of nodules in chest X-Rays.

Download Paper

AUTOMATING CARDIOMETRIC MEASUREMENTS IN CHEST X-RAYS

Published in 2021 International Conference "Nonlinearity, Information and Robotics" (NIR), 2020

The analysis of the positions, shapes, and sizes of thoracicorgans is an internationally established practice for radiolo-gists. The considerable amount of time spent on manual mea-surements of roentgenographic features reveals the need fora computerized approach for the automation of these mea-surements. In this work, we introduce a new way for theannotation of the chest x-ray data in thoracic ratios estima-tion. Using a manually annotated dataset, we developed adeep learning-based approach to infer three cardiometrics inchest radiographs, namely the Cardiothoracic Ratio, Lupi Co-efficient, and Moore Coefficient. The cardiometrics of inter-est are defined as ratios of line segments drawn over chestX-rays. We encoded the line segments with landmarks andapplied an hourglass model for landmark detection. To thebest of our knowledge, this is the first study aiming to estimatetwo out of three aforementioned cardiometrics. We comparedthe performance of the proposed solution with intraobservervariability of a radiologist using the test-retest strategy with aone-year break. We found out that human performance is notequally consistent across different measurements with morethan 20% difference in the F1-score metric

Extracting clinical information from chest x-ray reports: A case study for Russian language

Published in 2020 NIR Innopolis, 2020

—In this paper, we analyze possible approaches for diagnosis identification in Russian medical reports. Firstly, we introduce the main problems of raw Russian medical reports preprocessing. Secondly, focusing on the embedding extraction method, we analyzed several publicly available models and discovered that the use of BERT model is a promising instrument for this task. Performing the first attempt to build the NLP system for the Russian medical report classification based on the embeddings extraction method, we formulated the main weaknesses that limit the use of the existing publicly available Russian NLP models in the medical-text domain. Having no labeled data available, we evaluate each model visually, analyzing embeddings representation in 2D field retrieved by dimensionality reduction using t-SNE. We assume that a good model will be able to place reports that describe the same diagnosis close to each other, while moving reports with distinct diagnoses far from each other, forming clusters. Finally, we proposed several ways of possible future research that, as we believe, will improve the results achieved in this field so far.

Download Paper

Guiding evolutionary strategies by differentiable robot simulators

Published in arXiv preprint, 2021

The use of simulators in robotics research is widespread, underpinning the majority of recent advances in the field. There are now more options available to researchers than ever before, however navigating through the plethora of choices in search of the right simulator is often non-trivial. Depending on the field of research and the scenario to be simulated there will often be a range of suitable physics simulators from which it is difficult to ascertain the most relevant one. We have compiled a broad review of physics simulators for use within the major fields of robotics research. More specifically, we navigate through key sub-domains and discuss the features, benefits, applications and use-cases of the different simulators categorised by the respective research communities. Our review provides an extensive index of the leading physics simulators applicable to robotics researchers and aims to assist them in choosing the best simulator for their use case.

Gaze-based attention to improve the classification of lung diseases

Published in SPIE Medical Imaging 2022, 2022

Detection of lung diseases from chest X-rays has been of great interest from the research community during the last decade. Despite the existence of large annotated public databases, computer-aided diagnostic solutions still fail on challenging rare abnormality cases. In this study, we investigated the paradigm of combining the analysis of chest X-rays and physician gaze patterns during the analysis of these X-rays to improve the computerized diagnostic accuracy. Tobii Eye Tracker 4C has been mounted to a physician workstation and his eye movements were recorded during the analysis of 400 chest X-rays in two days of work. The X-rays have been sampled from CheXpert, RSNA, and SIIM-ACR public databases labeled with 14 different pathology types. The task was formulated as a binary classification problem. A ResNet34-based neural network has been trained to map the input chest X-ray with the output physician gaze map and binary pathology label. The proposed network improved the diagnostic accuracy to 0.714 of the area under receiving operator curve (AUC) from 0.681 AUC obtained for the same ResNet34 trained to generate binary pathology labels alone. The proposed study has demonstrated the potential benefits of using gaze information in computerized diagnostic solutions.

Multi-landmark environment analysis with reinforcement learning for pelvic abnormality detection and quantification

Published in Medical Image Analysis, 2022

Morphological abnormalities of the femoroacetabular (hip) joint are among the most common human musculoskeletal disorders and often develop asymptomatically at early easily treatable stages. In this paper, we propose an automated framework for landmark-based detection and quantification of hip abnormalities from magnetic resonance (MR) images. The framework relies on a novel idea of multi-landmark environment analysis with reinforcement learning. In particular, we merge the concepts of the graphical lasso and Morris sensitivity analysis with deep neural networks to quantitatively estimate the contribution of individual landmark and landmark subgroup locations to the other landmark locations. Convolutional neural networks for image segmentation are utilized to propose the initial landmark locations, and landmark detection is then formulated as a reinforcement learning (RL) problem, where each landmark-agent can adjust its position by observing the local MR image neighborhood and the locations of the most-contributive landmarks. The framework was validated on T1-, T2- and proton density-weighted MR images of 260 patients with the aim to measure the lateral center-edge angle (LCEA), femoral neck-shaft angle (NSA), and the anterior and posterior acetabular sector angles (AASA and PASA) of the hip, and derive the quantitative abnormality metrics from these angles. The framework was successfully tested using the UNet and feature pyramid network (FPN) segmentation architectures for landmark proposal generation, and the deep Q-network (DeepQN), deep deterministic policy gradient (DDPG), twin delayed deep deterministic policy gradient (TD3), and actor-critic policy gradient (A2C) RL networks for landmark position optimization. The resulting overall landmark detection error of 1.5 mm and angle measurement error of 1.4° indicates a superior performance in comparison to existing methods. Moreover, the automatically estimated abnormality labels were in 95% agreement with those generated by an expert radiologist.

AI-based analysis of radiologist’s eye movements for fatigue estimation: a pilot study on chest X-rays

Published in SPIE Medical Imaging 2022, 2022

Radiologist-AI interaction is a novel area of research of potentially great impact. It has been observed in the literature that the radiologists’ performance deteriorates towards the shift ends and there is a visual change in their gaze patterns. However, the quantitative features in these patterns that would be predictive of fatigue have not yet been discovered. A radiologist was recruited to read chest X-rays, while his eye movements were recorded. His fatigue was measured using the target concentration test and Stroop test having the number of analyzed X-rays being the reference fatigue metric. A framework with two convolutional neural networks based on UNet and ResNeXt50 architectures was developed for the segmentation of lung fields. This segmentation was used to analyze radiologist’s gaze patterns. With a correlation coefficient of 0.82, the eye gaze features extracted lung segmentation exhibited the strongest fatigue predictive powers in contrast to alternative features.

AI-based extraction of radiologists gaze patterns corresponding to lung regions

Published in Lecture Notes in Networks and Systems, 2022

The continuing growth of radiological examinations and the recent extreme workload on radiology divisions around the world make it necessary to optimize radiologist’s workflow. Intellectual radiologist’s gaze pattern analysis is a promising task in various applications of radiologist-AI interaction. In this paper, we propose an approach based on the Transformer deep learning architecture for predicting the current lung anatomical region by a short radiologist’s gaze history alone without any image information. For this study, we did a series of eye-tracking experiments with practicing radiologists and collected 400 chest X-ray images, that were analyzed independently by 4 doctors. From the results, we conclude the possibility of extracting useful information in various radiologist-AI interaction applications based only on a gaze history.

Artificial intelligence for the analysis of workload-related changes in radiologists’ gaze patterns

Published in IEEE Journal of Biomedical and Health Informatics, 2022

Around 60-80% of radiological errors are attributed to overlooked abnormalities, the rate of which increases at the end of work shifts. In this study, we run an experiment to investigate if artificial intelligence (AI) can assist in detecting radiologists’ gaze patterns that correlate with fatigue. A retrospective database of lung X-ray images with the reference diagnoses was used. The X-ray images were acquired from 400 subjects with a mean age of 49 ± 17, and 61% men. Four practicing radiologists read these images while their eye movements were recorded. The radiologists passed a series of concentration tests at prearranged breaks of the experiment. A U-Net neural network was adapted to annotate lung anatomy on X-rays and calculate coverage and information gain features from the radiologists’ eye movements over lung fields. The lung coverage, information gain, and eye tracker-based features were compared with the cumulative work done (CDW) label for each radiologist. The gaze-traveled distance, X-ray coverage, and lung coverage statistically significantly (p < 0.01) deteriorated with cumulative work done (CWD) for three out of four radiologists. The reading time and information gain over lungs statistically significantly deteriorated for all four radiologists. We discovered a novel AI-based metric blending reading time, speed, and organ coverage, which can be used to predict changes in the fatigue-related image reading patterns.

A 178-clinical-center experiment of integrating AI solutions for lung pathology diagnosis

Published in Scientific Reports, 2023

Abstract In 2020, an experiment testing AI solutions for lung X-ray analysis on a multi-hospital network was conducted. The multi-hospital network linked 178 Moscow state healthcare centers, where all chest X-rays from the network were redirected to a research facility, analyzed with AI, and returned to the centers. The experiment was formulated as a public competition with monetary awards for participating industrial and research teams. The task was to perform the binary detection of abnormalities from chest X-rays. For the objective real-life evaluation, no training X-rays were provided to the participants. This paper presents one of the top-performing AI frameworks from this experiment. First, the framework used two EfficientNets, histograms of gradients, Haar feature ensembles, and local binary patterns to recognize whether an input image represents an acceptable lung X-ray sample, meaning the X-ray is not grayscale inverted, is a frontal chest X-ray, and completely captures both lung fields. Second, the framework extracted the region with lung fields and then passed them to a multi-head DenseNet, where the heads recognized the patient’s gender, age and the potential presence of abnormalities, and generated the heatmap with the abnormality regions highlighted. During one month of the experiment from 11.23.2020 to 12.25.2020, 17,888 cases have been analyzed by the framework with 11,902 cases having radiological reports with the reference diagnoses that were unequivocally parsed by the experiment organizers. The performance measured in terms of the area under receiving operator curve (AUC) was 0.77. The AUC for individual diseases ranged from 0.55 for herniation to 0.90 for pneumothorax.

Anatomy-Preserving Counterfactual Edits in Breast MRI via Guided Diffusion

Published in Lecture Notes in Computer Science, 2025

Dynamic contrast-enhanced breast MRI is highly sensitive but difficult to interpret. We ask whether counterfactual edits produced by a latent diffusion model can reveal image cues linked to tumor presence while preserving anatomy. Starting from a latent diffusion backbone, we fine-tune the U-Net denoiser on the MAMA-MIA dataset and generate slice-level edits via DDIM inversion and null-text inversion. We study two operations: curtailment (suppression of tumor evidence) and exaggeration (amplification or addition of tumor-like features). Semantic impact is quantified with a frozen 3D nnU-Net trained on expert tumor masks. Across cohorts, curtailment yields a consistent 18–43% mean reduction in predicted tumor extent, whereas exaggeration is less reliable, highlighting an asymmetry between subtractive and additive edits. These results suggest diffusion-based counterfactuals provide interpretable, anatomy-preserving “what-if” views that complement saliency maps for model auditing, hypothesis generation, and training (Code and trained models are available at https://github.com/Luab/breast-mri-counterfactuals .)

Towards generating more interpretable counterfactuals via concept vectors: a preliminary study on chest X-rays

Published in Communications in Computer and Information Science, 2025

An essential step in deploying medical imaging models in real-world settings is ensuring that these models align with established clinical knowledge and produce outputs that are interpretable by clinicians. Recent research has demonstrated that the latent spaces of pre-trained models contain directions corresponding to high-level concepts. In this work, we focus on mapping clinical concepts into the latent space of generative models to identify corresponding Concept Activation Vectors (CAVs). Using a simple reconstruction autoencoder, we show that user-defined concepts can be linked to image-level features without requiring explicit training on class labels. We demonstrate that the concepts extracted using CAVs are stable across datasets. This indicates that in principle they can be used to generate visual explanations that highlight clinically relevant features. Our method produces counterfactual explanations by traversing the latent space in the direction of concept vector, allowing for the exaggeration or curtailment of specific clinical features. Our preliminary study on chest X-ray data shows that while the method performs well for larger pathologies such as cardiomegaly, challenges remain for smaller pathologies like atelectasis due to limitations in reconstruction fidelity and feature localization. Our investigation, although preliminary, demonstrates potential for producing interpretable, concept-based explanations in medical imaging, which is a key step towards improving trust in AI-driven diagnostics and foster adoption. Despite the approach does not beat the baseline in its current form, we believe that finding clinically relevant concepts in the latent space of pre-trained models provides a pathway for generating visual explanations that align with clinical knowledge. We envision taking this approach further by improving the underlying generative model and going beyond single vector by modeling concept subspace as a distribution.

ABRA: Agent Benchmark for Radiology Applications

Published in 2026 arXiv, 2026

Existing medical-agent benchmarks deliver imaging as pre-selected samples, never as an environment the agent must navigate. We introduce ABRA, a radiology-agent benchmark in which the agent operates an OHIF viewer and an Orthanc DICOM server through twenty-one function-calling tools that span slice navigation, windowing, series selection, pixel-coordinate annotation, and structured reporting. ABRA contains 655 programmatically generated tasks across three difficulty tiers and eight types (viewer control, metadata QA, vision probe, annotation, longitudinal comparison, BI-RADS reporting, and oracle variants of annotation and BI-RADS reporting), drawn from LIDC-IDRI, Duke Breast Cancer MRI, and NLST New-Lesion LongCT. Each episode is scored along Planning, Execution, and Outcome (Bluethgen et al., 2025) by task-type-specific automatic scorers. Ten current models, five closed-weight and five open-weight, reach at least 89% Execution on real annotation but only 0-25% Outcome; on the paired oracle variant where a simulated detector supplies the finding, Outcome on the same task reaches 69-100% across the models evaluated, localising the bottleneck to perception rather than tool orchestration. Code, task generators, and scorers are released at https://github.com/Luab/ABRA

Download Paper

talks

teaching