Publications

Poster Presentation

1. Towards AI based Freehand 3D Ultrasound Reconstruction , IEEE SAUS 2024, IIT Gandhinagar.

2. AI-Assisted Fetal Head Anatomical Plane Detection in Secure Ultrasound Reconstructed Volumetric Data, 3rd MS Divisional Research Symposium in "Applications of AI-ML in Engineering" at IISc Bangalore.

3.  Detection of Anatomical Structures of a Fetal Phantom in Ultrasound Using YOLO,  IISc-Siemens Workshop on Artificial Intelligence in Precision Medicine at IISc Bangalore.

4. AI-Assisted Fetal Head Anatomical Plane Detection for Remote Antenatal Screening. Ripples in Dept of Design and Manufacturing, IISc Bangalore.

Publications 

1. P. Saladi and M. Arora, "Anatomical Plane Classification for Fetal Head Ultrasound - Addressing Class Imbalance with Deep Learning," 2025 IEEE International Biomedical Instrumentation and Technology Conference (IBITeC), Yogyakarta, Indonesia, 2025, pp. 89-94, doi: 10.1109/IBITeC66306.2025.11473194.

Abstract: Ultrasound is the safest existing imaging modality to monitor fetal development. Reliable identification of the standard anatomical landmark planes of the fetal head is critical in obstetric ultrasound to ensure accurate biometric measurements and evaluations of fetal development. In this study, we propose a deep learning-based classification framework that uses transfer learning for feature extraction for automated detection of anatomical landmark head planes from ultrasound images followed by a classification block. The classification process consists of two stages: (1) Head vs. Not-a-Head classification and (2) identification of anatomical landmark head planes into trans-ventricular, trans-thalamic, trans-cerebellar or other. We evaluated multiple convolutional neural network (CNN) architectures such as VGG-16, MobileNet, MobileNetV2, EfficientNet-B0 and EfficientNet-B5, as feature extraction blocks, at different learning rates and analyzed performance in terms of overall accuracy, class-wise accuracy, precision, recall and F1-scores. Our findings demonstrate that lighter architectures such as MobileNet and MobileNetV2 with carefully tuned learning rate outperform deeper networks for this task, especially in imbalanced datasets. Additionally, we implemented a custom class-weighted loss function to address class imbalance, improving minority class detection. The best model was then deployed on a realtime ultrasound imaging system for classification to demonstrate its potential application for antenatal scanning. 

keywords: {Feeds;MIMICs;Millimeter wave integrated circuits;Monolithic integrated circuits;Protocols;Radio access networks;Regional area networks;Videos;HTTP;Video equipment;Fetal Ultrasound;Anatomical plane classifcation;Transfer learning;Deep learning;SecureUS;Maternal healthcare},URL: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11473194&isnumber=11472795

2. P. Saladi and Y. Kalepu, "Electromagnetic Inverse Scattering Problem Solved by DConvNet and Adapted Attention U-Net," 2023 IEEE 12th International Conference on Communication Systems and Network Technologies (CSNT), Bhopal, India, 2023, pp. 101-104, doi: 10.1109/CSNT57126.2023.10134734. 

Abstract: A two-stage Deep Learning strategy for resolving the electromagnetic inverse scattering problem (ISP) is proposed in this paper. The two steps involve utilizing Deep Convolutional Neural Network (DConvNet) to draw out features from the Scattered field data and refine the image reconstruction using adapted version of Attention U-Net. We observed that our proposed method delivered better results in terms of image quality and reconstruction precision. Moreover, we observed the outcomes of Attention U-Net, DConvNet, and DRCNN for ISP and found that Attention U-Net gave better image refinement.

3. Pravallika Saladi, Yaswanth Kalepu. Deep Learning Strategy for Resolving Electromagnetic Inverse Scattering Problem with Phase-less Data. Authorea. July 27, 2023. DOI: 10.22541/au.169045202.28967646/v1