[DSC DACH 25] From Data to Decisions: Building Responsible AI for the Future of Medical Imaging.pptx

Artificial Intelligence (AI) is transforming the healthcare sector by enhancing decision-making, improving patient outcomes, and optimizing diagnostic workflows. In particular, AI-driven data science is enabling clinicians to extract meaningful insights from vast volumes of medical imaging data—ranging from radiology and pathology scans to real-time imaging modalities—supporting more accurate and timely diagnoses. As these technologies become increasingly integral to clinical practice, it is essential to ensure they are developed and deployed responsibly. This talk will explore state-of-the-art applications of AI in medical imaging, including computer vision for image interpretation, anomaly detection, treatment predcition and workflow automation, through the lens of Responsible AI. This talk will also address critical ethical considerations such as bias mitigation, transparency, accountability, and data privacy, along with regulatory and implementation challenges. Attendees will gain insights into how responsible AI can bridge the gap between raw imaging data and actionable clinical intelligence, ultimately leading to more fair, effective, and patient-centered healthcare solutions.

1. From Data to Decisions: Building Responsible AI for the Future of Medical Imaging Ass.-Prof. Dr. Sepideh Hatamikia Assistant professor at Danube Private University (DPU), Krems, Austria, Head of CAROM research group (Clinical AI-Research in Omics and Medcial Data Sceince ), Deputy Scientific Director at ACMIT (Austrian Center for Medical Innovation and Technology) October 2025

2. 2 The healthcare industry generates massive volumes of data=‘Big Data’ Data science enables healthcare providers to derive meaningful insights from this wealth of information! Identify high-risk patients Optimizing the diagnosis Enhance preventive care strategies Optimizing treatment plans Enable efficient immediate decision-making Leading to improved patient outcomes Data Science and Analytics

3. 3 Medical Imaging in Breast Cancer Mammography – detects early tumors and microcalcifications . Ultrasound – distinguishes solid from cystic lesions , guides biopsies . MRI – identifies hidden or multifocal tumors , monitors treatment response . PET/CT – assesses spread and helps in staging . Histopathology – gold standard for confirming diagnosis and grading tumors. And m ore.. Scintimammography Mammography MRI Ultrasound CT PET Histopathology Microwave

4. AI can make several tasks from a radiologist automatic and even with higher accuracy and speed! but it can also do tasks which are not able to be done by a radiologists (main art of the AI) such as: Early detection of cancer years in advance Predict patient benefits from a specific type of treatment e.g., chemotherapy Image analysis by AI algorithm is superior to the traditional image analysis method. Large data sets of medical images can be processed to identify hidden and complex patterns and features that may not be seen by human eye! AI and Medical Imaging in breast cancer AI revolution in medical diagnostic! Advantage: Higher accuracy Higher speed Higher efficiency  

5. AI- based Prediction of Treatment Outcome Medical imaging data Breast cancer AI prediction model development Neoadjuvant chemotherapy (NAC) has been used increasingly in patients with breast cancer. However, prediction of response to NAT remain the great challenge! Prediction of which patient will respond and not can help avoiding toxicity (chemotherapy) to those who do not benefit from it!

6. 6 Hatamikia S , et al. Breast MRI radiomics and machine learning-based predictions of response to neoadjuvant chemotherapy - How are they affected by variations in tumor delineation ? Comput Struct Biotechnol J. 2023 Nov 19;23:52-63. doi : 10.1016/j.csbj.2023.11.016. PMID: 38125296; PMCID: PMC10730996. Radiomics and AI to predict response to neoadjuvant chemotherapy for breast cancer patients Radiomics: Extraction of quantitative imaging features using advanced feature analysis Prediction of Neoadjuvant Chemotherapy (NAC) Response, a Radiomics Approach

7. Responsible AI Challenges: AI models often perform well only on the data they were trained on High dependency on: Lesion segmentation quality Image acquisition and pre-processing methods The hospital or center where data were collected Performance drops significantly when : Different radiologists perform the segmentation Data come from another MRI device or medical center Radiomics research lacks standardization across studies and centers To ensure reliable clinical use , AI radiomics models must be generalizable and robust Generalization and robustness are essential for responsible AI in medicine

8. 8 Hatamikia S , et al. Breast MRI radiomics and machine learning-based predictions of response to neoadjuvant chemotherapy - How are they affected by variations in tumor delineation ? Comput Struct Biotechnol J. 2023 Nov 19;23:52-63. doi : 10.1016/j.csbj.2023.11.016. PMID: 38125296; PMCID: PMC10730996. Prediction of Neoadjuvant Chemotherapy (NAC) Response, a Radiomics Approach Dilation: grow VOI larger Erosion: shrink VOI Blue dots: original lesions Orange dots: after morphological operations Smoothing: smoothed VOI Randomization: randomized VOI Approximate ellipsoid Results : Best prediction with manually defined tumor VOIs: AUC 0.96 (HER2+) AUC 0.89 (TNBC) First comprehensive study on VOI modification effects inter- observer variability Key finding : VOI choice significantly impacts feature values , selection , and prediction performance Implication : Standardized tumor delineation is critical for reliable radiomics models Impact: Results can serve as a reference for future AI radiomics research

9. N4 bias field correction (BC) is a pre-processing algorithm used to correct for intensity inhomogeneities in MR images. Piecewise linear histogram equalization (PLHE) is a histogram-based normalization technique that is often used to enhance image contrast in MRI scans Prediction of Neoadjuvant Chemotherapy (NAC) Response, Influence of Different MR Normalization Algorithms

10. Evaluation & Key Findings We performed comprehensive evaluation across: Multiple datasets Various imaging device manufacturers Independent clinical centers Observed: Significant differences in MRI datasets and AI/radiomics features Results varied with different image normalization techniques Conclusion: Image normalization should be treated as a model hyperparameter in radiomics and AI research and standardization needs to be done for this step Prediction of Neoadjuvant Chemotherapy (NAC) Response, Influence of Different MR Normalization Algorithms

11. T1-weighted T1-weighted T2-weighted T2-weighted DCE DCE MRI acquisition Patient Pre-NAC (T0) Post-NAC (T3) Diagnosis of NAC Response U sing L ongitudinal MRI Images AI-based prediction of pCR surgery 11 To diagnose who receive the Pathologic complete response ( pCR ) Goal: to forgo the surgery for those who achieved pCR To predict who benefits from NAC Goal: to exclude those who dose not benefit

12. DCE Convnext Feature extraction Fusion Cross-attention pCR diagnosis Train and test on an open access data set on 22 different centers in USA T1-weighted T2-weighted DCE 12 Weighted Ensemble ( TabT 50% · FCNN 20% · XGB 30%) AUC: 0.91 F1: 0.77 Accuracy : 0.90 MRI sequences Diagnosis of NAC Response U sing L ongitudinal MRI Images

13. Medical Image Analysis and AI in Breast Diagnosis Ultrasound- based detection and malignancy prediction of breast lesions eligible for biopsy A multi-center clinical -scenario study using nomograms , large language models , and radiologist evaluation Purpose: Develop and externally validate integrated ultrasound nomograms ( a statistical tool that combines several clinical or imaging variables to predict the probability of a specific outcome for an individual patient) Combine BI-RADS ( established qualitative imaging features ) with quantitative morphometric characteristics Compare performance with : -Expert radiologists -Large language m odels (ChatGPT-o3 & o4-mini-high) - Evaluate performance in biopsy recommendation and malignancy prediction for breast lesions

14. Ultrasound- B ased D etection and M alignancy P rediction of B reast L esions E ligible for B iopsy M ulti-center, multi-national study 1,747 women with pathologically confirmed breast lesions Ultrasound data from three centers (Iran & Turkey)

15. Comparative evaluation : Three radiologists (1 senior , 2 general ) Two ChatGPT variants independently interpreted images Ultrasound- B ased D etection and M alignancy P rediction of B reast L esions E ligible for B iopsy

16. Ultrasound- B ased D etection and M alignancy P rediction of B reast L esions E ligible for B iopsy

17. Responsible AI Fairness and Equity in AI Ensure AI models perform reliably across diverse patient populations covering age , gender , ethnicity , geography , and hospital types Identify and minimize hidden biases that could lead to unequal or unsafe care Train models on multi-center, multi- vendor , and multi-population data to improve generalizability Increase robustness by accounting for observer variability in ground truth generation Standardize preprocessing steps across the development pipeline to reduce bias and variability

18. Responsible AI in medical imaging means: Fairness and Equity Transparency and Explainability Radiologists and clinicians need to understand why an AI model made such a decision, not just that it did. Responsible AI provides interpretable outputs heatmaps, confidence intervals, probability scores, or clear reasoning, so doctors can verify results instead of blindly trusting them. 3. Validation and Safety Rigorous testing before clinical use: across multiple hospitals, imaging devices, and patient groups. Meeting regulatory standards (e.g. FDA, CE mark) with clear evidence of benefit and risk mitigation. Continuous monitoring after deployment , because data shifts over time (change of demography, change of medical devices etc ) and models can degrade. Responsible AI

19. 4. Integration into Clinical Workflow Tools should support radiologists, not burden them with extra clicks or isolated dashboards. Responsible AI fits seamlessly into PACS or EHR systems and enhances efficiency instead of disrupting care. 5. Accountability and Governance Clear ownership: Who is responsible if the AI makes an error? The developer, the hospital, the clinician? Transparent reporting of limitations, no overhyping capabilities. Mechanisms for recall, retraining, or updating models when problems are detected. 6. Privacy and Security Protecting patient data during training and deployment. Ensuring compliance with data protection laws (GDPR, HIPAA). Responsible AI

20. Conclusion From Data to Decisions, Toward Responsible AI in Medicine AI is transforming medical imaging with higher accuracy, speed, and efficiency . Radiomics and deep learning unlock hidden patterns beyond human perception. However, responsibility, fairness, and transparency are essential for safe clinical use. Standardization and validation across centers ensure robustness and trust. The future lies in collaboration between clinicians, data scientists, and regulators to build AI that truly supports better patient outcomes. → Responsible AI = Reliable Medicine

21. Acknowledgment

22. Thanks for your attention

23. References https://www.techscience.com/iasc/v37n1/52678/html https://radiologica.com.tr/en/artificial-intelligence-and-mammography Who’s Responsible When AI Fails in Healthcare? - Respocare Insights

24. Who’s Responsible When AI Makes a Mistake in Healthcare? Responsibility for proper use and maintenance of device remains with the providers . In case of assistive Al, the physician remains fully liable .