Swathi Gowroju
Swathi Gowroju has a total of 27 co-authors. Here are the top co-authors by number of shared publications:
Shilpa Choudhary
Neil Gogte Institute of Technology, India
3 shared publications- Devineni Gireesh Kumar
Department of CSE, Symbiosis Institute of Technology, Pune, Pune, India
2 shared publications - Sandeep Kumar
Symbiosis Institute of Technology Nagpur Campus, Symbiosis International (Deemed University) Pune India
2 shared publications - Aarti
Lovely Professional University, India
1 shared publication - Anshu Ghimire
Department of Computer Science and Engineering, Nepal Engineering College, Kathmandu, Nepal
1 shared publication - Arun Singh
Lovely Professional University
1 shared publication
Latest Publications
- Open Access
- 3 February 2026
- Published by Intelligence Science and Technology Press Inc. in Journal of Artificial Intelligence and Technology
Validated Swin Transformer-Based Deep Learning Pipeline with Cross-Validation and McNemar’s Test for Multi-Class Hemorrhage Classification in Traumatic Brain Injury CT Scans
Traumatic brain injury (TBI) remains a major global health concern, where rapid and accurate identification of intracranial hemorrhage on computed tomography (CT) scans is essential for improving patient outcomes. Manual radiological assessment, although clinically effective, is time-consuming and subject to inter-observer variability, creating a need for reliable automated diagnostic systems. Most existing artificial intelligence (AI) studies focus on binary hemorrhage detection or limited subtype classification, leaving a gap in robust multi-class intracranial hemorrhage categorization. This study presents a validated deep learning pipeline based on the Swin Transformer—a hierarchical Vision Transformer architecture that employs windowed and shifted multi-head self-attention to capture both local and global contextual information. A balanced dataset of 60,000 CT images across six clinically relevant classes was generated through structured augmentation of a publicly available source, ensuring fairness and improved generalization. The model was optimized using AdamW, cosine annealing scheduling, and label smoothing to enhance stability. The proposed framework achieved a classification accuracy of 82.02%, outperforming the strong baseline EfficientNetV2-S model (80.17%). Statistical significance was established using McNemar’s test (p < 0.0001), and robustness was demonstrated through 5-fold cross-validation. Gradient-weighted Class Activation Mapping (Grad-CAM)++- based interpretability analysis further confirmed that the model consistently highlighted clinically meaningful hemorrhagic regions, reinforcing its diagnostic relevance. These findings establish the Swin Transformer as an effective, interpretable, and statistically validated solution for multi-class intracranial hemorrhage classification, with strong potential to support radiologists in emergency triage and improve clinical decision-making.
- 3 February 2026
- Published by Intelligence Science and Technology Press Inc. in Journal of Artificial Intelligence and Technology
Open AccessValidated Swin Transformer-Based Deep Learning Pipeline with Cross-Validation and McNemar’s Test for Multi-Class Hemorrhage Classification in Traumatic Brain Injury CT Scans
Traumatic brain injury (TBI) remains a major global health concern, where rapid and accurate identification of intracranial hemorrhage on computed tomography (CT) scans is essential for improving patient outcomes. Manual radiological assessment, although clinically effective, is time-consuming and subject to inter-observer variability, creating a need for reliable automated diagnostic systems. Most existing artificial intelligence (AI) studies focus on binary hemorrhage detection or limited subtype classification, leaving a gap in robust multi-class intracranial hemorrhage categorization. This study presents a validated deep learning pipeline based on the Swin Transformer—a hierarchical Vision Transformer architecture that employs windowed and shifted multi-head self-attention to capture both local and global contextual information. A balanced dataset of 60,000 CT images across six clinically relevant classes was generated through structured augmentation of a publicly available source, ensuring fairness and improved generalization. The model was optimized using AdamW, cosine annealing scheduling, and label smoothing to enhance stability. The proposed framework achieved a classification accuracy of 82.02%, outperforming the strong baseline EfficientNetV2-S model (80.17%). Statistical significance was established using McNemar’s test (p < 0.0001), and robustness was demonstrated through 5-fold cross-validation. Gradient-weighted Class Activation Mapping (Grad-CAM)++- based interpretability analysis further confirmed that the model consistently highlighted clinically meaningful hemorrhagic regions, reinforcing its diagnostic relevance. These findings establish the Swin Transformer as an effective, interpretable, and statistically validated solution for multi-class intracranial hemorrhage classification, with strong potential to support radiologists in emergency triage and improve clinical decision-making.
- 1 January 2026
- Published by GN1 Sistemas e Publicacoes Ltd. in ITEGAM-Journal of Engineering and Technology for Industrial Applications (ITEGAM-JETIA)
- Vol. 12 (59) , 1882-1892
Automated Lung Nodule Segmentation Using an MH-UNet Deep Learning Framework
One of the leading causes of abnormal deaths globally is lung cancer, and pertaining to the improvement of patient's prognosis depends directly on early identification. To diagnose such condition the proposed system has identified lung nodule segmentation from CT scan pictures that is robust and effective. The goal of this methodology is to segregate lung nodules from CT scan pictures using a deep learning network called MH-Interposed system using the MH-UNet deep learning model to separate lung nodules from CT scan pictures. To predict lung cancer, the model was trained to identify region of interest of lungs and detect nodules. Proposed system uses 888 CT scan pictures and their corresponding ground truth nodule masks to perform segmentation. Jaccard Index and Dice Score were calculated to evaluate the performance of the system. The model detected with great accuracy in lung nodule detection with an average Dice Score of 0.92 and a Jaccard Index of 0.85.
- 1 January 2026
- Published by GN1 Sistemas e Publicacoes Ltd. in ITEGAM-Journal of Engineering and Technology for Industrial Applications (ITEGAM-JETIA)
- Vol. 12 (59) , 1882-1892
Automated Lung Nodule Segmentation Using an MH-UNet Deep Learning Framework
One of the leading causes of abnormal deaths globally is lung cancer, and pertaining to the improvement of patient's prognosis depends directly on early identification. To diagnose such condition the proposed system has identified lung nodule segmentation from CT scan pictures that is robust and effective. The goal of this methodology is to segregate lung nodules from CT scan pictures using a deep learning network called MH-Interposed system using the MH-UNet deep learning model to separate lung nodules from CT scan pictures. To predict lung cancer, the model was trained to identify region of interest of lungs and detect nodules. Proposed system uses 888 CT scan pictures and their corresponding ground truth nodule masks to perform segmentation. Jaccard Index and Dice Score were calculated to evaluate the performance of the system. The model detected with great accuracy in lung nodule detection with an average Dice Score of 0.92 and a Jaccard Index of 0.85.
- 1 January 2026
- Published by GN1 Sistemas e Publicacoes Ltd. in ITEGAM-Journal of Engineering and Technology for Industrial Applications (ITEGAM-JETIA)
- Vol. 12 (59) , 1857-1870
A Deep Dive into Style-Based Age Manipulation: Revisiting SAM for Lifespan Face Editing
Change of facial age aims to realistically alter a person's photograph's age without loss of identity, a very important aspect for forensic work, entertainment, and all digital media. The Style-Based Age Manipulation (SAM) model is based on the expressive latent space of StyleGAN2 to enable photorealistic, seamless age progression and regression across the human life cycle (5 to 100 years). SAM uses a pSp encoder to embed real images into the 𝑊 + latent space and a new ageing encoder for age-conditioned, non-linear transformations with a fixed StyleGAN2 generator for high-quality synthesis. 𝐿𝐿2, LPIPS, ArcFace, age regression, and cycle consistency losses have been optimized in SAM to improve visual quality, identity fidelity, and reversible transformations. Its novelty lies in a regression-based architecture that enables continuous, disentangled age control, in contrast to previous methods that rely on discrete age bins or linear latent traversals. SAM outperforms on LPIPS (mean: 0.11), ArcFace similarity (mean: 0.93), and age classification accuracy (91% ± 5 years). Qualitative results provide lifelike smooth age transitions, while quantitative measures show 10% improved identity preservation and 18% better visual fidelity compared with CNN baselines. A strong, generalizable method is developed in SAM and sets a new standard for identity-holding facial age change, with broad practical applicability.
- 1 January 2026
- Published by GN1 Sistemas e Publicacoes Ltd. in ITEGAM-Journal of Engineering and Technology for Industrial Applications (ITEGAM-JETIA)
- Vol. 12 (59) , 1857-1870
A Deep Dive into Style-Based Age Manipulation: Revisiting SAM for Lifespan Face Editing
Change of facial age aims to realistically alter a person's photograph's age without loss of identity, a very important aspect for forensic work, entertainment, and all digital media. The Style-Based Age Manipulation (SAM) model is based on the expressive latent space of StyleGAN2 to enable photorealistic, seamless age progression and regression across the human life cycle (5 to 100 years). SAM uses a pSp encoder to embed real images into the 𝑊 + latent space and a new ageing encoder for age-conditioned, non-linear transformations with a fixed StyleGAN2 generator for high-quality synthesis. 𝐿𝐿2, LPIPS, ArcFace, age regression, and cycle consistency losses have been optimized in SAM to improve visual quality, identity fidelity, and reversible transformations. Its novelty lies in a regression-based architecture that enables continuous, disentangled age control, in contrast to previous methods that rely on discrete age bins or linear latent traversals. SAM outperforms on LPIPS (mean: 0.11), ArcFace similarity (mean: 0.93), and age classification accuracy (91% ± 5 years). Qualitative results provide lifelike smooth age transitions, while quantitative measures show 10% improved identity preservation and 18% better visual fidelity compared with CNN baselines. A strong, generalizable method is developed in SAM and sets a new standard for identity-holding facial age change, with broad practical applicability.
- 21 February 2025
- Published by IGI Global Scientific Publishing
- p. 259-282
Visual Speech Recognition With Viseme-Phoneme Mapping Improving Lip-Reading Systems Using GAN
Lip reading is to deduce the meaning of speech from the movements of the lips. Lip reading has advanced significantly as a result of the introduction of large-scale datasets and the advancement of deep learning technologies. Analyzing visual signals from a speaker's lips movements to comprehend spoken words is a method called lip-reading, or visual speech recognition. In order to simplify the recognition process, the suggested method uses a viseme-to-phoneme mapping technique, which divides visually identical phonemes into separate visemes. The Proposed method takes input of mouth images and analyzes it to derive characteristics that match different visemes. To precisely link visemes to their corresponding phonemes or words, these attributes are examined using sophisticated machine learning models, such as convolutional and recurrent neural networks. The efficacy of the system is assessed using key performance measures such as Word Error Rate (WER), accuracy, and Visual Speech Intelligibility (VSI). With a WER of 25%, accuracy of 85%, and VSI score of 0.72, the findings show that the suggested viseme-to-phoneme mapping greatly improves the accuracy and resilience of lip-reading models. The hearing handicapped can benefit greatly from this method, which can also improve voice recognition in loud or silent contexts and human-computer interaction. However, as many sounds have very similar or identical lip forms, it can be challenging to discriminate between visually similar phonemes when lip-reading without aural information.
- 21 February 2025
- Published by IGI Global Scientific Publishing
- p. 259-282
Visual Speech Recognition With Viseme-Phoneme Mapping Improving Lip-Reading Systems Using GAN
Lip reading is to deduce the meaning of speech from the movements of the lips. Lip reading has advanced significantly as a result of the introduction of large-scale datasets and the advancement of deep learning technologies. Analyzing visual signals from a speaker's lips movements to comprehend spoken words is a method called lip-reading, or visual speech recognition. In order to simplify the recognition process, the suggested method uses a viseme-to-phoneme mapping technique, which divides visually identical phonemes into separate visemes. The Proposed method takes input of mouth images and analyzes it to derive characteristics that match different visemes. To precisely link visemes to their corresponding phonemes or words, these attributes are examined using sophisticated machine learning models, such as convolutional and recurrent neural networks. The efficacy of the system is assessed using key performance measures such as Word Error Rate (WER), accuracy, and Visual Speech Intelligibility (VSI). With a WER of 25%, accuracy of 85%, and VSI score of 0.72, the findings show that the suggested viseme-to-phoneme mapping greatly improves the accuracy and resilience of lip-reading models. The hearing handicapped can benefit greatly from this method, which can also improve voice recognition in loud or silent contexts and human-computer interaction. However, as many sounds have very similar or identical lip forms, it can be challenging to discriminate between visually similar phonemes when lip-reading without aural information.
- 21 February 2025
- Published by IGI Global Scientific Publishing
- p. 347-370
Eye-Tracking Control System for Intuitive Hands-Free Human-Computer Interaction and Mouse Navigation
With today's sophisticated technologies, there is always room for advancement in the computing industry. As of right now, HCI is in demand since it helps people with quadriplegia. This study presents an effective Human-computer interaction (HCI) solution for amputees and individuals with hand impairments. The developed system translates eye motions, such as winking, into facial movements, such as expanding the lips and rotating the face (pitch and yaw) according to the cursor's path on the mouse. If the amputee can discern the angle at which the tip of their nose is pointing and detect eye movement, they can freely move the pointer by tracking the motion of their nose and eye to the mouse. A “tracking mouse” will be very helpful to an amputee. “The system under discussion uses a simple webcam and requires facial recognition using ML languages and a few more software packages. Building the face detector using a sliding window method, a linear classifier, and the Histogram of Oriented Gradients feature, there is no need for additional gear or sensors, and it is hands-free.
- 21 February 2025
- Published by IGI Global Scientific Publishing
- p. 347-370
Eye-Tracking Control System for Intuitive Hands-Free Human-Computer Interaction and Mouse Navigation
With today's sophisticated technologies, there is always room for advancement in the computing industry. As of right now, HCI is in demand since it helps people with quadriplegia. This study presents an effective Human-computer interaction (HCI) solution for amputees and individuals with hand impairments. The developed system translates eye motions, such as winking, into facial movements, such as expanding the lips and rotating the face (pitch and yaw) according to the cursor's path on the mouse. If the amputee can discern the angle at which the tip of their nose is pointing and detect eye movement, they can freely move the pointer by tracking the motion of their nose and eye to the mouse. A “tracking mouse” will be very helpful to an amputee. “The system under discussion uses a simple webcam and requires facial recognition using ML languages and a few more software packages. Building the face detector using a sliding window method, a linear classifier, and the Histogram of Oriented Gradients feature, there is no need for additional gear or sensors, and it is hands-free.
- 29 November 2024
- Published by IGI Global Scientific Publishing
- p. 429-458
Leveraging Neuromorphic Computing for Human Action Detection With Deep Neural Networks
NMC(Neuro Morphic Computing) simulates and mimics the process of human brain operations using artificial neurons. The creation of human actions can be used to test and improve neuromorphic models, which are simulations of complex behaviours and interactions found in biological creatures. It can be contributed to the development of decision making capabilities in robots over time. The proposed system hence aims to use various deep learning models that train on UCF101 dataset to predict the set of actions or actions from given video clip. The proposed work aims to construct deep learning model, trained on large-scale datasets like UCF101, to efficiently learn and represent complex patterns and characteristics in human behaviors. Through the use of these models in the framework of neuromorphic computing, scientists can enhance the potential of artificial neural networks to more precisely describe and comprehend human behavior.
- 29 November 2024
- Published by IGI Global Scientific Publishing
- p. 429-458
Leveraging Neuromorphic Computing for Human Action Detection With Deep Neural Networks
NMC(Neuro Morphic Computing) simulates and mimics the process of human brain operations using artificial neurons. The creation of human actions can be used to test and improve neuromorphic models, which are simulations of complex behaviours and interactions found in biological creatures. It can be contributed to the development of decision making capabilities in robots over time. The proposed system hence aims to use various deep learning models that train on UCF101 dataset to predict the set of actions or actions from given video clip. The proposed work aims to construct deep learning model, trained on large-scale datasets like UCF101, to efficiently learn and represent complex patterns and characteristics in human behaviors. Through the use of these models in the framework of neuromorphic computing, scientists can enhance the potential of artificial neural networks to more precisely describe and comprehend human behavior.
- 29 November 2024
- Published by IGI Global Scientific Publishing
- p. 261-288
Classification of Moderate and Advanced Dementia Patients Using Gradient Boosting Machine Technique
In the twenty-first century, caring for persons with Dementia's has become extremely difficult due to the prevalence of dementia cases. Using data from the OASIS (Open Access Series of Imaging Studies) program provided by the University of Washington Dementia's Disease Research Center, the study presents a new predictive model for Dementia's. Dementia, a chronic condition and it's become a serious health concern in adults. Various methods of data imputation, preprocessing, and transformation were used to prepare the data for model training. Machine learning algorithms, including AdaBoost (AB), Decision Tree (DT), Exclusion Tree (ET), Gradient Boost (GB), K-Nearest Neighbor (KNN), Logistic Regression (LR), Naive Bayes (NB;, Random Forest (RF), and Support Vector Machine (SVM), were used in this field. These algorithms were evaluated on both the complete feature set and a subset of features selected via the Least Absolute Shrinkage and Selection Operator (LASSO) method. Comparative analysis based on accuracy, precision, and other metrics showed that the proposed method achieved the highest accuracy of 96.77% using Support Vector Machine (SVM) with all feature sets, further refined and applied, has great potential for the diagnosis of early Dementia's disease (AD) disease.
- 29 November 2024
- Published by IGI Global Scientific Publishing
- p. 261-288
Classification of Moderate and Advanced Dementia Patients Using Gradient Boosting Machine Technique
In the twenty-first century, caring for persons with Dementia's has become extremely difficult due to the prevalence of dementia cases. Using data from the OASIS (Open Access Series of Imaging Studies) program provided by the University of Washington Dementia's Disease Research Center, the study presents a new predictive model for Dementia's. Dementia, a chronic condition and it's become a serious health concern in adults. Various methods of data imputation, preprocessing, and transformation were used to prepare the data for model training. Machine learning algorithms, including AdaBoost (AB), Decision Tree (DT), Exclusion Tree (ET), Gradient Boost (GB), K-Nearest Neighbor (KNN), Logistic Regression (LR), Naive Bayes (NB;, Random Forest (RF), and Support Vector Machine (SVM), were used in this field. These algorithms were evaluated on both the complete feature set and a subset of features selected via the Least Absolute Shrinkage and Selection Operator (LASSO) method. Comparative analysis based on accuracy, precision, and other metrics showed that the proposed method achieved the highest accuracy of 96.77% using Support Vector Machine (SVM) with all feature sets, further refined and applied, has great potential for the diagnosis of early Dementia's disease (AD) disease.
- Open Access
- 13 October 2022
- Published by Hindawi Limited in Mathematical Problems in Engineering
- Vol. 2022, 1-24
Deep Neural Network for Accurate Age Group Prediction through Pupil Using the Optimized UNet Model
Predicting age automatically from the image is a difficult task and shortening the challenge to be more concise is also a challenging task. Nevertheless, the existing implementations using manually designed features using a wide variety of input features using benchmark datasets are unsatisfactory as they suffer from unknown subject information. It is challenging to judge CNN’s performance using such approaches. The proposed system performs the segmentation through UNet without using a dense layer to perform the segmentation and classification. The proposed system uses the skip connection to hold the loss at the max-pooling layer. Also, the morphological processing and probabilistic classification served as the proposed system’s novelty. The proposed method used three benchmark datasets, MMU, CASIA, and UBIRIS, to experiment with building a training model and tested using various optimization techniques to perform an accurate segmentation. To further test and improve the quality of the proposed method, we experimented with random images. The proposed system’s accuracy is 96% when experimented on random images of subjects collected purely for experimentation. Three optimizers, namely, Stochastic Gradient Descent, RMS Prop, and Adaptive Moment Optimizer, were experimented with in the proposed system to fit the system. The average accuracy we received using optimizers is 71.9, 84.3, and 96.0 for the loss value of 2.36, 2.30, and 1.82, respectively.
- 13 October 2022
- Published by Hindawi Limited in Mathematical Problems in Engineering
- Vol. 2022, 1-24
Open AccessDeep Neural Network for Accurate Age Group Prediction through Pupil Using the Optimized UNet Model
Predicting age automatically from the image is a difficult task and shortening the challenge to be more concise is also a challenging task. Nevertheless, the existing implementations using manually designed features using a wide variety of input features using benchmark datasets are unsatisfactory as they suffer from unknown subject information. It is challenging to judge CNN’s performance using such approaches. The proposed system performs the segmentation through UNet without using a dense layer to perform the segmentation and classification. The proposed system uses the skip connection to hold the loss at the max-pooling layer. Also, the morphological processing and probabilistic classification served as the proposed system’s novelty. The proposed method used three benchmark datasets, MMU, CASIA, and UBIRIS, to experiment with building a training model and tested using various optimization techniques to perform an accurate segmentation. To further test and improve the quality of the proposed method, we experimented with random images. The proposed system’s accuracy is 96% when experimented on random images of subjects collected purely for experimentation. Three optimizers, namely, Stochastic Gradient Descent, RMS Prop, and Adaptive Moment Optimizer, were experimented with in the proposed system to fit the system. The average accuracy we received using optimizers is 71.9, 84.3, and 96.0 for the loss value of 2.36, 2.30, and 1.82, respectively.