India AI Impact Summit Insights
All sessions

AI for Medical Imaging and Diagnostics

Contents

Executive Summary

This panel discussion brings together leading clinicians, AI engineers, and pathologists from India's premier medical institutions to examine AI's transformative role in healthcare diagnostics. The panelists emphasize that AI functions most effectively as a force multiplier and clinical decision-support tool—dramatically improving efficiency, accuracy, and access to care—while stressing that human clinicians must remain central to diagnosis and treatment decisions. The discussion highlights India-specific challenges: the need for locally-trained models using diverse Indian populations, data standardization across institutions, infrastructure deployment to rural areas, and the cultural shift required to build trust in AI systems.

Key Takeaways

  1. AI is a Clinical Companion, Not a Replacement: The future is "humanized AI"—systems that augment physician decision-making, handle routine tasks, and flag urgent cases for expert review. The doctor remains accountable and in the decision loop.

  2. Local Data Pools Beat Global Models: India cannot rely on algorithms trained on Western populations. Institutions across east, west, north, and south India must standardize data formats and share anonymized datasets to build population-specific models that actually work for Indian patients.

  3. Implementation is Harder Than Innovation: The technical barriers are solvable; the real challenges are data standardization, interoperability between hospitals, clinician training, workflow redesign, and cultural acceptance. Without addressing these, excellent AI systems sit unused.

  4. Trust Requires Transparency: Clinicians will adopt AI only when they understand how it works. Explainable AI methods (attention maps, feature importance, SHAP values) that show which image regions or data points drove a diagnosis are essential for acceptance.

  5. Rural and Affordable Solutions Need Dedicated Development: Government-backed initiatives (like Rashni for cataracts or thermal imaging for breast cancer) provide templates. Scaling requires partnerships between startups, government, and academic institutions to develop cost-effective, simple-to-use tools for primary care settings.

Key Topics Covered

  • AI as a Clinical Tool: How AI enhances diagnostic accuracy, reduces time-to-diagnosis, and supports clinical decision-making
  • Early Detection and Screening: AI's potential to identify diseases at earlier, more treatable stages
  • Disease-Specific Applications: Cancer imaging (prostate, breast, gallbladder, kidney, colorectal, lung), chronic pain diagnosis, pathology, diabetic complications, and ophthalmology
  • Trust and Explainability: Building physician confidence through explainable AI (XAI), local validation, and continuous monitoring
  • Data Privacy and Interoperability: Managing patient data, genomic privacy, and integration of disparate hospital systems
  • Rural and Underserved Healthcare: Deployment strategies for tier-2 and tier-3 cities, cost considerations, and workforce readiness
  • Institutional Barriers: Training clinicians, adapting workflows, and breaking down silos between hospitals
  • Data Standardization: The critical need for consistent data formats across India to enable population-scale AI models

Key Points & Insights

  1. AI as a "Force Multiplier": Rather than replacing doctors, AI amplifies physician capacity—enabling one radiologist to effectively review 20-25 patients instead of 10, while maintaining or improving diagnostic quality. Time-to-diagnosis can be cut from hours to minutes.

  2. Triaging and Prioritization: AI excels at rapidly screening high-volume cases (e.g., 50 gallbladder specimens daily) to identify which cases require expert attention, allowing pathologists and radiologists to focus their expertise where it matters most.

  3. Reduced Interobserver Variability: In histopathology, AI-standardized analysis reduces the subjective differences between experts viewing the same slide, functioning like a "second expert" and improving consistency in diagnosis.

  4. Predictive Diagnostics Over Reactive: Medicine is shifting from "What do you have?" to "What will you have?"—AI enables detection of disease markers years before symptoms manifest, exemplified by models predicting diabetic complications or early malignancy onset.

  5. India Needs India-Specific Models: Western-trained AI models often underperform on Indian populations due to genetic diversity, dietary differences, disease prevalence patterns, and symptom presentations. Local validation and locally-trained models are non-negotiable.

  6. Data is the Bottleneck: The primary challenges are not algorithmic but logistical—institutions lack standardized, digitized, interoperable data. Hard copies and proprietary file formats prevent pooling datasets needed for robust population-scale models.

  7. Explainability is Essential for Trust: The shift from "black box" to "explainable AI" (XAI) using attention mechanisms and feature importance visualization allows clinicians to understand why an AI system makes a recommendation, not just what it recommends.

  8. Institutional Readiness is Low: Many hospitals lack the infrastructure, staff training, and workflow redesign needed to integrate AI effectively. Staff retention concerns, acceptance issues, and the perception that "machines are replacing doctors" hinder adoption.

  9. Cost and Government Support: Deployment to rural Primary Health Centers (PHCs) and Community Health Centers (CHCs) requires government funding and simplified, low-cost solutions. One example: the Rashni app for cataract screening in Uttar Pradesh.

  10. Data Governance Complexity: Organizations deploying AI must negotiate data privacy (especially genomic data), IP protection, federated learning architectures, and regulatory compliance while still enabling algorithm improvement through data sharing.

Notable Quotes or Statements

  • Dr. Sujit Gautam: "AI offers individualized treatment—which particular patient is going to have more optimum relief and positive aspects of therapy."

  • Dr. Chandan: "For us, AI is a force multiplier. It not only helps in our diagnosis but it saves a lot of time." (Flagging abnormal cases among 300-400 daily X-rays, so doctors focus on the abnormal 20%.)

  • Dr. Kundan Singh Jofal: "We cut short the patient assessment time from 3 days to 1 day by using AI to predict treatment eligibility—this has real impact on patient care."

  • Dr. Alok Sharma: "AI reduces interobserver variability. It's like having a second expert looking at the case simultaneously, and it does mundane work like counting mitosis, freeing us for more fruitful work."

  • Dr. Chufal: "There are three important things AI is helping us with: seeing better, deciding better, and working faster. ... We cannot give everything to AI to decide; the doctor must stay in the loop."

  • Nidhi (AI Engineer): "The problem is AI is a black box. But with explainable AI in the last 3-4 years, the black box is becoming gray and will become white. It's not just engineers—doctors can now understand the reasoning."

  • Dr. Sujit: "The AI tool output should be explainable. The doctor should know how the output was generated. If we don't understand, some mishap might happen."

  • Nidhi: "AI is an important term. Anything intelligent cannot be artificial. It's 80 billion mathematical operations and differential calculus. The real question isn't whether doctors will be there—it's whether AI can help doctors save more lives."

  • Dr. Chandan: "AI models need to be humanized and supervised—patients should not feel 'a machine is answering, the doctor is not helping me.'"

Speakers & Organizations Mentioned

Panel Members (Identified)

  • MK Dutta — Director, Center for Artificial Intelligence, AMIT University; Computer scientist specializing in healthcare AI and imaging biomarkers
  • Dr. Sujit Gautam — Professor, Department of Anesthesiology, SGPGIMS (Sanjay Gandhi Postgraduate Institute of Medical Sciences); researching AI for chronic pain diagnosis
  • Prof. Nidhi Goyel — Professor, Department of Electronics and Communication Engineering, Indira Gandhi Delhi Technical University; expertise in deep learning for medical imaging (e.g., wireless capsule endoscopy)
  • Dr. Chandan — Professor of Radiology, Department of Radio-Diagnosis & Intervention Radiology, New Delhi; working on cancer imaging (gallbladder, kidney, prostate, colorectal cancers)
  • Dr. Kundan Singh Jofal — Chief, Department of Radiotherapy Services, Rajiv Gandhi Cancer Institute; focus on thoracic and sarcoma/lymphoma malignancies; 40% of publications are AI-based
  • Dr. Alok Sharma — Head, R&D Services, Department of Pathology & Electron Microscopy, Dr. Agarwal's (largest diagnostic provider in India); working on AI for routine diagnostics

Institutions & Organizations

  • AMIT University, Noida
  • SGPGIMS (Sanjay Gandhi Postgraduate Institute of Medical Sciences)
  • Indira Gandhi Delhi Technical University (IGDTU)
  • Rajiv Gandhi Cancer Institute
  • Dr. Agarwal's Diagnostic Services (largest diagnostic provider in India)
  • ICMR (Indian Council of Medical Research) — funding AI projects for diagnostics
  • NITI Aayog (National Institution for Transforming India)
  • Government of Uttar Pradesh (Rashni cataract screening app)
  • Max Healthcare (mentioned via questioner Dr. Hinata Bhya U)
  • Sapien Labs Diagnostics (SL Diagnostics)

Other Referenced Entities

  • King George Medical University (where panelist guides a PhD candidate)
  • RML Hospital, Delhi
  • Gangaram Hospital, Delhi
  • Google (Sundar Pichai's investments in diabetic retinopathy and cataract detection)
  • NITI Aayog AI initiatives

Technical Concepts & Resources

AI/ML Methodologies

  • Explainable AI (XAI): Attention mechanisms, feature importance plots, SHAP values, visualization of decision drivers
  • Federated Learning: Distributed model training without centralizing sensitive data
  • Retrieval-Augmented Generation (RAG): Vector databases paired with language models for privacy-preserving inference
  • Deep Learning Models: Convolutional neural networks for image classification; mentioned for X-ray, CT, ultrasound, MRI analysis
  • Predictive Modeling: Risk score development (cardiac scores, metabolic complication predictions)
  • Triaging Algorithms: Automated prioritization of high-risk cases from high-volume screening

Medical Imaging & Diagnostic Applications

  • Radiography & CT Imaging: Lung nodule detection, gallbladder cancer, kidney cancer, prostate cancer, breast cancer screening
  • Ultrasound: Gallbladder cancer detection, obstetric imaging, abdominal pain assessment
  • MRI: Prostate cancer grading and stratification
  • Wireless Capsule Endoscopy: Gastroenterology; reducing 1 lakh frames of 4-hour review to 20 frames in 2-3 minutes
  • Histopathology & Whole Slide Imaging (WSI): Digital pathology for cancer grading, mitosis counting, standardized analysis
  • Thermography: Infrared imaging for breast cancer screening (e.g., ThermaLogics)
  • Ophthalmology: Diabetic retinopathy detection; cataract screening (pupillary images)

Data & Infrastructure Concepts

  • Whole Slide Imaging (WSI): Digital scanning of histology slides enabling AI analysis
  • Data Standardization: Consistent electronic formats across institutions for disease documentation
  • Interoperability: Integration of multiple hospital systems under unified interface
  • Data Drift: Monitoring output changes when data distribution shifts (e.g., temporal, geographic, demographic)
  • Inter-Observer Variability: Subjective differences between expert assessments (reduced by AI)

Specific Algorithms/Tools/Datasets

  • Rashni App — Government of Uttar Pradesh cataract screening tool using pupillary images
  • Thermalogics — Thermography-based breast cancer detection device
  • ICMR Database (56 Lakh Patients) — Large Indian patient cohort for virus diagnostics research (referenced by MK Dutta)
  • Google's Diabetic Retinopathy/Cataract Detection Models — Large-scale screening initiatives
  • Chart GPT — Referenced as example of AI doctors are not yet ready to trust for diagnosis

Key Risk Scoring Systems Mentioned

  • Cardiac risk scores
  • Prostate cancer risk scores (PI-RADS)
  • Diabetes complication prediction (diabetic nephropathy, metabolic complications)
  • Cancer prognostication scores

Challenges & Research Areas

  • Data Privacy in Genomics: Managing DNA data from Indian populations without exposing to external parties
  • Model Generalization: Validating Western-trained models on Indian population; need for diversity (geographic, age, gender, socioeconomic)
  • Black Box to White Box Conversion: Transitioning from unexplainable predictions to interpretable models
  • Cold Chain Management Analogy: Applying vaccine distribution logistics to AI deployment in rural settings

Additional Context

Geographic & Public Health Focus

  • Emphasis on North-South-East-West diversity in India: disease prevalence, dietary habits (e.g., soft foods in South leading to different dental problems), genetic variation
  • Gallbladder Cancer Belt: High-risk region along Gangatic plains; low prevalence in southern India
  • Tier-2 and Tier-3 Cities: Shortage of specialists; opportunity for AI-enabled remote diagnostics
  • PHC/CHC Deployment: Scaling to Primary Health Centers and Community Health Centers with ASHA (Accredited Social Health Activists) workers

Regulatory & Governance Issues

  • Data Governance: Balancing privacy (HIPAA-equivalent), IP protection, and collaborative data sharing for public health benefit
  • Government Funding: ICMR projects, NITI Aayog initiatives as funding sources for AI healthcare research

Cultural & Organizational Barriers

  • Staff retention concerns ("I'm retiring in 6 months—why learn this?")
  • Patient perception: Technology perceived as replacing, not assisting, physicians
  • Clinician-Engineer communication gap: Different vocabularies and objectives
  • Workflow integration: Adding AI alongside existing systems vs. genuine system redesign