Prof. Bapi Raju explores the healthcare landscape against the backdrop of AI evolution, discussing key challenges and future directions while emphasizing the importance of ethical AI deployment and human-AI collaboration.

Artificial Intelligence (AI) has been a driving force in technological advancements across industries, with healthcare being a signifi cant benefi ciary. Initially, AI in healthcare focused on discriminative models for classifi cation tasks, such as anomaly detection in medical images and disease progression prediction from patient records. The advent of deep learning introduced complex models like Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), and Gated Recurrent Units GRUs for text-based applications, while architectures like Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Deep Convolutional GANs (DCGANs) enabled synthetic medical image generation. Table 1 gives a taxonomy of tasks and models that span the Discriminative AI revolution that happened over the last 2 decades.

Note: SSD: Single-Shot Detector, consists of a classification backbone + SSD head (another CNN): a method for detecting objects in images using a single deep neural network; GBM: Gradient Boosting Machine; XGBoost: Extreme Gradient Boosting; Faster (Region-based CNN) R-CNN predicts object class and bounding boxes, while Mask R-CNN is an extension that adds pixel-level segmentation, both are object detection algorithms that use regions to localize objects in an image; FCN: Fully Convolutional Network, type of CNN used for image segmentation; TFT: Temporal Fusion Transformer

GenAI Beyond Conventional Applications
GenAI extends traditional AI, in particular, Discriminative AI, by generating new medical content, rather than just analyzing existing data. Key applications include:
Text Generation: Automated radiology report writing, chatbot-based patient interactions, and AI-assisted medical literature summarization.
Image Generation: AI-assisted medical imaging enhancements, synthetic data generation for rare disease research, and real-time image refi nement.
Audio Generation: AI-powered voice synthesis for medical dictation, patient communication aids, and vocal biomarker analysis for disease detection.
Video Generation: AI-driven surgical simulation, automated medical training videos, and deep learning-based anomaly detection in real-time medical scans.

Key Pillars of GenAI
Successful development and deployment of a GenAI solution rests on 8 pillars (see Fig. 1), namely, the deep learning (DL) models; availability of large-scale datasets; access to computational power; design of appropriate training techniques to the DL models; evaluation metrics to measure the quality, creativity and diversity; applications and use cases; human in the loop (HITL) to facilitate collaboration, fi ne-tuning, and lastly framing and adopting ethical and responsible practices.

GenAI in Medical Imaging and Diagnostics
One of the most promising applications of GenAI is in medical imaging and diagnostics. Models like CheXagent interpret chest X-rays, providing automated insights into conditions like tuberculosis and pneumonia1. The Medical SAM2 Model enhances 3D abdominal segmentation, crucial for precise organ localization2. The advent of foundation models has opened unprecedented opportunities of building task-agnostic general models that can be fi ne-tuned for various downstream tasks3. Foundation models trained on vast multimodal datasets, such as BiomedGPT and Med-Gemini signifi cantly improve medical imaging accuracy and report generation4,5. In Fig. 2, some results from CheXagent with Indian data are depicted6. Results from the experiments using Indian data on foundation models trained on North

Key Pillars of GenAI
Successful development and deployment of a GenAI solution rests on 8 pillars (see Fig. 1), namely, the deep learning (DL) models; availability of large-scale datasets; access to computational power; design of appropriate training techniques to the DL models; evaluation metrics to measure the quality, creativity and diversity; applications and use cases; human in the loop (HITL) to facilitate collaboration, fine-tuning, and lastly framing and adopting ethical and responsible practices.

American samples clearly show the challenges that we would encounter.

Challenges and Future Directions
Despite its potential, GenAI faces challenges in healthcare, including image quality control and regulatory constraints. Ensuring that AI-generated images and interpretations are clinically valid requires rigorous validation processes. In diverse healthcare contexts, building robust AI models faces challenges like lack of standardized datasets and varying imaging equipment. Particularly in India, the urban versus rural divide needs to be bridged, in the sense that medical data is concentrated in urban hospitals, but the rural areas have limited digital records. If not addressed, such skewness can potentially introduce urban bias into models. Large repositories from North America such as MIMIC dataset or a UK Biobank equivalent in India is possible with the recent eff orts of IndiaAI mission in creating the compute, ecosystem, and datasets. The ethical guidelines formulated by ICMR for application of AI in biomedical research and healthcare are timely7.

The future of GenAI in healthcare lies in augmenting clinical decision-making, enhancing diagnostic accuracy, and automating tedious tasks. Successful integration requires addressing technical challenges like bias in AI models and ensuring human-AI collaboration in clinical settings.

References
1. Chen, Zhihong, et al. (2024). CheXagent: Towards a foundation model for chest x-ray interpretation. arXivpreprint arXiv:2401.12208
2. Zhu, Jiayuan, Qi, Yunli, and Wu, Junde (2024). Medical sam 2: Segment medical images as video via segment anything model 2. arXiv preprint arXiv:2408.00874
3. Bommasani, Rishi, et al. (2021). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258
4. Zhang, K., Zhou, R., Adhikarla, E. et al. (2024). A generalist vision–language foundation model for diverse biomedical tasks. Nat Med 30, 3129–3141. https://doi.org/10.1038/s41591-024-03185-2
5. Yang, Lin, et al. (2024). Advancing multimodal medical capabilities of Gemini. arXiv preprint arXiv:2405.03162
6. Chest X-rays Tuberculosis data from India: https://www.kaggle.com/datasets/raddar/chest-xrays-tuberculosis-from-india
7. ICMR (2023). Ethical guidelines for application of Artificial Intelligence in Biomedical Research and Healthcare, ISBN: 978-93-5811-343-3

Conclusion
Generative AI represents a new frontier in healthcare innovation, off ering unprecedented possibilities in medical imaging, diagnostics, and automated clinical workflows. By addressing key obstacles and ensuring responsible AI adoption, we can fully harness the power of GenAI to improve patient outcomes and revolutionize modern medicine.

This article was initially published in the February ’25 edition of TechForward Dispatch