Publications

Development of a deep learning model for intussusception using point-of-care ultrasound

Published in Frontiers in Radiology, 2026

This study evaluated the feasibility of using deep learning to detect intussusception from point-of-care ultrasound (POCUS) images in pediatric patients. A dataset comprising 785 ultrasound images and video clips from 49 patients, including 8 confirmed cases of intussusception, was preprocessed by enhancing image quality, extracting the region of interest, and converting video clips into static image frames, resulting in 1,582 intussusception and 1,965 normal images for model training. Multiple artificial intelligence approaches, including classical machine learning, ensemble learning, transfer learning, and fine-tuned deep learning models, were developed and compared. The fine-tuned models demonstrated superior performance across patient-level, image-level, and threshold-based analyses, with the threshold-based approach achieving the highest predictive accuracy. These findings demonstrate the feasibility of developing an AI-assisted screening tool for identifying the ultrasound target sign of intussusception using relatively small POCUS datasets, highlighting its potential to improve early diagnosis in emergency settings where pediatric imaging expertise may be limited.

Recommended citation: Anand Thyagachandran, Lefchak B, Murthy HA, Bergmann KR and Madhok M. "Development of a deep learning model for intussusception using point-of-care ultrasound". in Front. Radiol. 6:1812587. doi: 10.3389/fradi.2026.1812587 https://www.frontiersin.org/journals/radiology/articles/10.3389/fradi.2026.1812587/full

Breast Cancer Segmentation using UNet and Global Convolutional Networks

Published in International Conference on Pattern Recognition (ICPR), 2024

Breast ultrasound (BUS) imaging techniques have become efficient tools for cancer diagnosis. Convolutional neural network (CNN) based encoder-decoder architectures have been widely used for the automated segmentation of tumours in BUS images, assisting in breast cancer diagnoses. However, these models have limitations in capturing long-range dependencies. To overcome this limitation, various deep learning techniques, such as atrous convolution, attention mechanisms, and transformer encoder-based models, have been introduced to capture long-range dependencies in feature maps, improving segmentation accuracy by considering larger receptive fields and global context. As modelling techniques evolve, there is a shift towards more complex and intricate designs. This study proposes a simple yet effective model that combines UNet and Global Convolutional Network (GCN) architectures for breast lesion segmentation. By leveraging the GCN block, our model captures broader receptive fields with a simpler design strategy. We have demonstrated the efficacy of our approach through various experiments, including kernel size analysis, model component evaluation, and data preprocessing assessment. The proposed model has been evaluated using four-fold cross-validation with BUSI and Dataset-B datasets. Additionally, models trained on both datasets have been validated with a blind test dataset, where our model demonstrates better performance compared to state-of-the-art methods, achieving a 4.9% and 6.7% improvement in Intersection over Union (IoU) score, respectively. The robustness analysis and external validation experiments underscore the superior generalization performance of our model in breast lesion segmentation tasks.

Recommended citation: A. Thyagachandran, A Ahmed, "Breast Cancer Segmentation using UNet and Global Convolutional Networks," 2024, International Conference on Pattern Recognition (ICPR) Kolkata, India, 2024.

Linear Prediction on Cent Scale for Fundamental~Frequency~(f0) Analysis

Published in Proceedings of JASA Express Letters, 2024

Understanding the fundamental frequency and harmonic content of audio signals is crucial for many applications in music analysis, including music transcription, audio synthesis, and genre identification. This study formulates a signal processing approach combining Linear Prediction (LP) analysis and the Cent scale to characterize audio signals’ pitch and harmonic structure accurately. Pitch tracking on the LP spectrum in the Cent scale provides more accurate and reliable pitch estimation, especially in the presence of noise or overlapping harmonics. The Cent scale aligns the harmonics of different notes more closely, making it easier to discern the correct pitch.

Recommended citation: Gowriprasad R, Anand T, R Aravind, and Hema A Murthy 2024) Linear Prediction on Cent Scale for Fundamental~Frequency~(f0) Analysis Proc. JASA Express Letters 2024

Identification and Severity Assessment of COVID-19 Using Lung CT Scans

Published in IEEE Access, 2023

The COVID-19 pandemic, caused by the SARS-CoV-2 virus, continues to have a significant impact on the global population. To effectively triage patients and understand the progression of the disease, a metric-based analysis of diagnostic techniques is necessary. The objective of the present study is to identify COVID-19 from chest CT scans and determine the extent of severity, defined by a severity score that indicates the volume of infection. An unsupervised preprocessing pipeline is proposed to extract relevant clinical features and utilize this information to employ a pretrained ImageNet EfficientNetB5 model to extract discriminative features. Subsequently, a shallow feed-forward neural network is trained to classify the CT scans into three classes, namely COVID-19, Community-Acquired Pneumonia, and Normal. Through various ablation studies, we find that a domain-specific preprocessing pipeline has a significant positive impact on classification accuracy. The infection segmentation mask generated from the preprocessed pipeline performs better than state-of-the-art supervised semantic segmentation models. Further, the estimated infection severity score is observed to be well correlated with radiologists’ assessments. The results confirm the importance of domain-specific preprocessing for training machine learning algorithms.

Recommended citation: A. Thyagachandran, A. Balachandran and H. A. Murthy, "Identification and Severity Assessment of COVID-19 Using Lung CT Scans," in IEEE Access, vol. 11, pp. 124542-124555, 2023, doi: 10.1109/ACCESS.2023.3330238 https://ieeexplore.ieee.org/document/10309131

Ensemble Methods For Enhanced Covid-19 CT Scan Severity Analysis

Published in IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), 2023

Computed Tomography (CT) scans provide a high-resolution image of the lungs, allowing clinicians to identify the severity of infections in COVID-19 patients. This paper presents a domain knowledge-based pipeline for extracting infection regions from COVID-19 patients using a combination of image-processing algorithms and a pre-trained UNET model. Then, an infection rate-based feature vector is generated for each CT scan. The infection severity is then classified into four categories using an ensemble of three machine-learning models: Random Forest, Support Vector Machines, and Extremely Randomized Trees. The proposed system is evaluated on the validation and test datasets with a macro F1 score of 58% and 46.31%, respectively. Our proposed model has achieved 3rd place in the severity detection challenge as part of the IEEE ICASSP 2023: AI-enabled Medical Image Analysis Workshop and COVID-19 Diagnosis Competition (AI-MIACOV19D).

Recommended citation: A. Thyagachandran and H. A. Murthy, "Ensemble Methods For Enhanced Covid-19 CT Scan Severity Analysis," 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), Rhodes Island, Greece, 2023, pp. 1-5, doi: 10.1109/ICASSPW59220.2023.10193538 https://ieeexplore.ieee.org/document/10193538

Dual Script E2E Framework for Multilingual and Code-Switching ASR

Published in Proceedings of Interspeech, 2021

India is home to multiple languages, and training automatic speech recognition (ASR) systems is challenging. Over time, each language has adopted words from other languages, such as English, leading to code-mixing. Most Indian languages also have their own unique scripts, which poses a major limitation in training multilingual and code-switching ASR systems. Inspired by results in text-to-speech synthesis, in this paper, we use an in-house rule-based phoneme-level common label set (CLS) representation to train multilingual and code-switching ASR for Indian languages. We propose two end-to-end (E2E) ASR systems. In the first system, the E2E model is trained on the CLS representation, and we use a novel data-driven backend to recover the native language script. In the second system, we propose a modification to the E2E model, wherein the CLS representation and the native language characters are used simultaneously for training. We show our results on the multilingual and code-switching (MUCS) ASR challenge 2021. Our best results achieve ≈6% and 5% improvement in word error rate over the baseline system for the multilingual and code-switching tasks, respectively, on the challenge development data.

Recommended citation: Kumar, M.G., Kuriakose, J., Thyagachandran, A., A, A.K., Seth, A., Prasad, L.V.S.V.D., Jaiswal, S., Prakash, A., Murthy, H.A. (2021) Dual Script E2E Framework for Multilingual and Code-Switching ASR. Proc. Interspeech 2021, 2441-2445, doi: 10.21437/Interspeech.2021-978 https://www.isca-archive.org/interspeech_2021/kumar21e_interspeech.pdf

EEG Responses for Dhrupad Stimulus: A Case Study

Published in International Conference on Signal Processing and Communications (SPCOM), 2020

Music possesses the capacity to affect the human brain deeply, weaving together elements like melody, rhythm, timbre, lyrics, and pitch to elicit a wide range of emotional and cognitive responses. This study delves into the impact of Dhrupad, a genre within Indian classical music known for its meditative qualities, on the EEG responses. The EEG data were collected while participants immersed themselves in serene melodies of a Dhrupad alāp performance on the Rudra Veena - a choice made for its gradual and soothing nature. Using 128-channel EEG, the signals from different brain regions are analysed. Multitaper spectrograms are employed to analyse the evolution of mental states. The proposed research reveals consistent patterns in the EEG signals for different levels of attentiveness. Especially, patterns correlate across different attentive subjects and suggest that the responses may be related to cognition.

Recommended citation: A. Thyagachandran, G. R, S. Jaiswal, R. Aravind and H. A. Murthy, "EEG Responses for Dhrupad Stimulus: A Case Study," 2024 International Conference on Signal Processing and Communications (SPCOM), Bangalore, India, 2024, pp. 1-5, doi: 10.1109/SPCOM60851.2024.10631650. (https://ieeexplore.ieee.org/abstract/document/10631650)

Seizure Detection Using Time Delay Neural Networks and LSTMs

Published in IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2020

Automatic detection of seizures from EEG signals is an important problem of interest for clinical institutions. EEG is a temporal signal collected from multiple spatial sources around the scalp. Efficient modelling of both temporal and spatial information is important to identify the seizures using EEG. In this paper, we propose a neural network system using the time-delay neural network to model temporal information (TDNN) and long short-term memory (LSTM) layer to model spatial information. On the development subset of the Temple University seizure dataset, the proposed system achieved a sensitivity of 23.32 % with 11.13 false alarms in 24 hours.

Recommended citation: A. Thyagachandran, M. Kumar, M. Sur, R. Aghoram and H. Murthy, "Seizure Detection Using Time Delay Neural Networks and LSTMs," 2020 IEEE Signal Processing in Medicine and Biology Symposium (SPMB), Philadelphia, PA, USA, 2020, pp. 1-5, doi: 10.1109/SPMB50085.2020.9353636. https://ieeexplore.ieee.org/abstract/document/9353636