Move Away HIPAA And GDPR, Here Comes CrypTFlow – Secure AI Inferencing Without Data Sharing

Move Away HIPAA And GDPR, Here Comes CrypTFlow – Secure AI Inferencing Without Data Sharing


  • Currently, running Artificial Intelligence (AI) algorithms on medical images requires either the sharing of medical images with developers of the algorithms, or sharing of the algorithms with the hospitals. Both these options are sub-optimal since there is always a real risk of patient privacy breach or of intellectual property theft.
  • Encryption is the process of converting data into a “secret code” using a “key” making the data meaningless for anyone without the key. The challenge is that, with current technology, the key needs to be shared with the AI developer, so that the data can be converted to its meaningful form, thereby compromising the security and privacy of the data.
  • We propose using CrypTFlow, which uses Multi-Party Computation and encryption to run AI algorithms on medical images without sharing the encryption key described above. This means that the images remain in the hospital network, the AI algorithm remains in the AI developer’s network, but the AI is still able to run on the images.
  • We will present the results of our experiments of running CheXpert, an AI algorithm, on Chest X-Rays.


  • Current privacy and intellectual property concerns with deploying AI algorithms in clinical practice • What is encryption? • What is multi-party computation (MPC)?
  • What is CrypTFlow and how can it help run AI algorithms without requiring data to be shared with AI developers?
  • Results of running CheXpert AI algorithm using CrypTFlow – accuracy, time and computation – The Future of secure AI deployment

The poster can be viewed here: CrypTFlow-Secure-AI-Inferencing

Automatic pre-population of normal chest x-ray reports using a high-sensitivity deep learning algorithm: a prospective study of clinical AI deployment (RPS1005b)


To evaluate a high-sensitivity deep learning algorithm for normal/abnormal chest x-ray (CXR) classification by deploying it in a real clinical setting.

Methods and materials:

A commercially available deep learning algorithm (QXR,, India) was integrated into the clinical workflow for a period of 3 months at an outpatient imaging facility. The algorithm, deployed on-premise, was integrated with PACS and RIS such that it automatically analysed all adult CXRs and reports for those which were determined to be “normal” were automatically populated in the RIS using HL7 messaging. Radiologists reviewed the CXRs as part of their regular workflow and ‘accepted’ or changed the pre-populated reports. Changes in reports were divided into ‘clinically insignificant’ and ‘clinically significant’ following which those CXRs with clinically significant changes were reviewed by a specialist chest radiologist with 8 years’ experience.


A total of 1,970 adult CXRs were analysed by AI, out of which 388 (19.69%) were identified to be normal. 361/388 (93.04%) of these were accepted by radiologists and in 14/388 (3.60%) clinically less significant changes (e.g. increased broncho-vascular markings) were made in reports. Upon review of the balance 13/388 (3.35%) CXRs, it was found that 12 had truly clinically significant missed findings by AI, including 3 with opacities, 3 with lymphadenopathy, 3 with blunted CP angle, 2 with nodules, and 1 with consolidation.


This study shows that there is a great potential to automate the identification of normal CXRs to a great degree, with very high sensitivity.

Validation of a high precision semantic search tool using a curated dataset containing related and unrelated reports of clinically relevant search terms (RPS 1005b)


To validate a sematic search tool by testing the search results for complex terms.

Methods and materials:

The tool consists of two pipelines: an offline indexing pipeline and a querying pipeline. The raw text from both reports and queries were first passed through a set of pre-processing steps; sentence tokenisation, spelling correction, negation detection, and word sense disambiguation. It was transformed into a concept plane followed by indexing or querying. During querying, additional concepts were added using a query expansion technique to include nearby related concepts. The validation was done on a set of 30 search queries, carefully curated by two radiologists. The reports that are related to the search queries were randomly selected with the help of keyword search and the text was re-read to determine its suitability to the queries. These reports formed the \”related\” group. Similarly, the reports that were not exactly satisfying the context of the search queries were categorised as the \”not related\” group. A set of 5 search queries and 250 reports were used for tuning the model initially. A total of 500 reports of the 10 search queries formed the corpus of the test set. The search results for each test query were evaluated and appropriate statistical analysis was performed.


The average precision and recall rates on 10 unseen queries on a small corpus for respective queries containing related and unrelated reports were 0.54 and 0.42. On a larger corpus containing 60 K reports, the average precision for these 15 queries was 0.6.


We describe a method to clinically validate a sematic search tool with high precision.

Next Generation Sequencing for the Practicing Radiologist

Next Generation Sequencing For The Practicing Radiologist


Next Generation Sequencing (NGS) is a technique of reading the entire genome or screening few relevant genes for aberrations. NGS sequences DNA parallelly thus resulting in reduced price and high speed of genome sequencing.

For Onco-Radiologists: NGS, can detect inherited or sporadic germline mutations or acquired changes i.e. somatic mutations on tumour tissues. Identification of germline mutations helps in confirming the diagnosis better management of diseases, prognosis and suggest preventive and surveillance measures for at risk unaffected relatives. Few radiological outcomes are spot diagnosis for a few cancer genetics syndrome. Somatic analysis of tumour biopsy helps in predicting the response of targeted therapy, prognostication and pharmacogenomics. For Ultrasonologists – For prenatal cases with abnormal USG findings, Non – Invasive Prenatal Techniques (NIPT), is a test that detects small amounts of cell free fetal DNA in the maternal bloodstream, allowing prenatal genetic diagnosis through maternal blood.


What is next generation sequencing?

Germline vs somatic mutation detection

Germline – applications in clinical care

Somatic mutations – large gene panels – what is their use?

NIPT – how can radiologists use it Imaging + genomics – the future?

Circulating Tumor DNA (ctDNA) – A Potential Adjunct to FDG-PET Imaging in Cancer Follow Up

Circulating Tumor DNA (CtDNA) – A Potential Adjunct To FDG-PET Imaging In Cancer Follow Up


• With the continuous advances in healthcare – Next Generation Sequencing and PET-CT are revolutionising cancer diagnostics

• Survival outcomes in cancer can be improved by early detection of metastatic disease or relapse during follow up

• Next Generation sequencing is an extremely high end modality wherein hundreds to millions of DNA molecules are sequenced in parallel

• CtDNA is an application of NGS to monitor the amount of tumor specific DNA in the peripheral blood of patient

• Using ctDNA, a non-invasive investigation, to monitor the remission status can potentially reduce the frequency of imaging required at follow up, hence reducing the side effects and discomfort

• According to Wong R et al, for colorectal cancer patients with indeterminate findings on routine investigations, ctDNA detection increases the probability that the findings indicate metastatic disease, including in a nonpredefined subset that also underwent FDG-PET imaging

• Hence, using ctDNA as an adjunct to FDG-PET imaging during cancer follow-up seems to be an acceptable proposition.


• What is ctDNA? How is it measured?

• What are the current guidelines for cancer follow up?

• Review of literature • Advantages and limitations of FDG-PET imaging

• Advantages and limitations of ctDNA

• Conclusion

Initial experiences with Amide Proton Transfer Imaging – Correlation between quantitative MTRasym ratio, rCBV and rCBF values in brain tumours

Poster Presentation at the European Congress of Radiology, Vienna, 2019


Amide proton transfer (APT) imaging is a new molecular MRI technique based on chemical exchange saturation transfer mechanism with proposed use as an imaging biomarker in brain tumours. We evaluated the correlation between Asymmetric magnetization transfer ratio (MTRasym) measured on APT scans with relative cerebral blood volume (rCBV) and relative CBF (rCBF) values measured on Dynamic susceptibility contrast (DSC) MR perfusion in patients with brain tumours.

Methods and Materials

Eleven patients underwent the clinical MRI protocol on 3.0 T MRI (Ingenia, Philips), followed by Perfusion and APT scans. Perfusion scans were performed using first-pass dynamic susceptibility-weighted contrast-enhanced (DSC) perfusion-weighted images and APT scans were performed by using a 2-dimensional single-slice radiofrequency-spoiled GRE protocol, with 21 frequency offsets from +5 to −5 ppm. ROIs were placed in solid parts of tumours on both perfusion and APT maps and the corresponding MTRasym, rCBV and rCBF values were measured.


The MTRasym, rCBV and rCBF values were plotted in each case and compared. The Pearson product-moment correlation coefficients were calculated between MTRasym and rCBV as well as MTRasym and rCBF were 0.3022 and 0.1158 suggesting a very weak correlation.


This study revealed a very weak correlation between quantitative values of APT scans and rCBV values measured on DSC perfusion scans. Although technically a positive correlation, the relationship between them is weak. Further prospective evaluation of APT is needed to corroborate our early observations.

T2 Nerve Imaging of the Brachial Plexus Using Compressed-SENSE Effect on Image Quality and Acquisition Time

(RSNA 2018, Wed Nov 28 2018 3:40PM – 3:50PM ROOM Z17)


In this study, compressed sensing is combined with the parallel-imaging or SENSE infrastructure, i.e., Compressed-SENSE (CSENSE), for accelerating anatomical MR data acquisition by exploiting the multi-element receiver coil sensitivity variation and sparsity constraining. We quantitatively evaluate the dual role of CSENSE imaging in reducing scan time without loss in resolution and in improving resolution with minimal increase in scan time.


Ten healthy volunteers were scanned on a 3.0T MRI (Ingenia, Philips) using a proprietary “3D NerveView Sequence” (NVS), a T2W TSE isotropic sequence (TR 2200, TE 170, 2 mm slice thickness). Three versions of NVS were developed, SEQ1 (no CSENSE), SEQ2 (CSENSE factor = 9) and SEQ3 (CSENSE factor = 9) with acquisition times 6:16, 3:22 and 8:19 minutes respectively. SEQ1 and SEQ2 had the same acquisition and recon matrix of 252X325 and 640X640 while SEQ3 had a higher resolution having acquisition and recon matrix of 316X414 and 720X720 respectively. Contrast-to-Noise Ratio (CNR) was measured in all at the levels of the nerve roots (C5 to T1 levels), trunks and cords using Shinkei’s formula of CNR = (SI-Nerve – SI-Muscle)/ (SI-Nerve + SI-Muscle) where SI is average Signal Intensity in the region.


There was no significant difference in CNR for roots in all three sequences, i.e. SEQ1 – 0.7102±0.102, SEQ2 – 0.7040±0.044 and SEQ3 – 0.7253±0.035. In the trunks, SEQ3 performed as well as SEQ1 with a CNR of 0.567±0.10 against 0.5497±0.09. SEQ2 had a lower CNR of 0.4843±0.11. At the level of cords, SEQ1 outperformed both SEQ2 and SEQ3 with a CNR of 0.4248±0.17 against 0.3079±0.11 and 0.3505±0.19 respectively. We note that CSENSE performs better in areas where average CNR is on the higher side.


While CSENSE gives radiologists flexibility of reducing time or increasing resolution, the decision of when and how to use CSENSE depends on the clinical context. It can be used to reduce scan time when root lesions are suspected and improve resolution when lesions of the trunks and cords are suspected.


Compressed Sensing for nerve imaging should be part of every radiologist and technologists’ arsenal to conduct patient-specific personalised scanning. Lower scan time leads to improved patient comfort and hence less motion artefacts, and higher resolution improves diagnostic accuracy of the scan.

Acceleration of MR Imaging of Spine Using Compressed-SENSE: A Comparison with Existing Standard of Care Clinical Acquisition Methods

(RSNA 2018, Wed Nov 28 2018 11:30AM – 11:40AM ROOM Z41)


To objectively evaluate the impact of Compressed-SENSE (CSENSE), a novel acquisition technique that combines compressed sensing with parallel imaging (or SENSE), on acquisition time and image quality in MR imaging of the spine.


Current standard of care clinical axial T1 and T2-weighted acquisitions of the cervical and lumbo-sacral spine were modified to obtain higher acceleration with CSENSE (CSENSE factor 1.3999). Twenty-six patients were scanned both, with and without CSENSE, on a 3.0 T wide-bore MRI (Ingenia, Philips Healthcare). The images were anonymised and shared with three specialist MRI radiologists blinded to the acquisition type. Both sets of images were rated on a scale of 1 to 5 for image quality and delineation of specific structures including vertebral bodies, lateral recess, neural foramina, facet joints, cauda equina, nerve roots, spinal cord and paraspinal muscles. Disc abnormalities, foraminal stenosis, nerve root compression and facet joint degeneration were also rated. Dicom metadata was analysed to assess scan time acceleration (ratio of original time to CSENSE time).


Interrater agreement on image quality was compared between the Normal and CSENSE scans via a multivariate non-parametric Hotelling\’s T2 test. There was no statistically significant difference between Normal and CSENSE scans in the 12 measures evaluated by each radiologist (alpha < 0.001 level). Further, the mean percentage observed agreement between Normal and CSENSE scans for all three measures across the radiologists was 93%. The average time acceleration achieved using CSENSE on axial T1 was 1.41 and on axial T2 was 1.36 and the average time saved was 58 seconds (29%) and 50 seconds (26%) respectively.


There is no difference in image quality between current standard of care and CSENSE-based T1 and T2 axial MRI scans of the spine. Compressed-SENSE in the spine can reliably replace current axial T1 and T2 acquisitions without loss in image quality and with significant reduction in scan time.


The potential for CSENSE to accelerate MRI acquisition without hampering image quality will increase patient throughput and patient compliance in MR scanning.

Amide Proton Transfer Imaging for Brain Tumors: Initial Experience and Current Status

(Educational Exhibit, RSNA 2018)
V Mahajan, MBBS, MBA; R Bhattacharjee, MENG; I Saha, PhD; H Mahajan, MD; S Panwar, MD; V Venugopal, MD


• Amide Proton Transfer (APT) Images contain quantitative colour maps that show protein / peptide content.
• APT imaging is able to accurately predict tumour grade in pre-surgical cases.
• APT imaging is showing promise in differentiating tumour recurrence from radiation necrosis.
• APT imaging can be used in combination with Susceptibility Weighted Imaging (SWI) based ITSS (Intra-Tumoral Susceptibility Score) to predict tumour grade when compared to Contrast-Enhanced Perfusion MRI.
• Some fundamental problems with the APT acquisition need to be addressed for it to come into day-to-day clinical practice, especially with respect to quantification and artefacts.


• Review of current brain tumour MR imaging protocols
• Mechanism behind APT imaging
• Review of current clinical literature on APT imaging for brain tumors
• APT for tumour recurrence vs radiation necrosis
• APT in glioma grading
• APT with SWI-based ITSS as a non-contrast replacement for CE-Perfusion MR
• Limitations and future directions for APT imaging in the brain