About
About this work
This site is a companion to a doctoral dissertation. Every claim here links to the number behind it, the sample it came from, and the limit on what it means.
Candidate
Binesh Sadanandan
Doctor of Philosophy in Engineering and Applied Science
University of New Haven, West Haven, Connecticut
Expected August 2026
Research in the Secure and Assured Intelligent Learning (SAIL) Lab, University of New Haven.
Committee
- Vahid Behzadan, Ph.D.
- Committee Chair and Advisor, Associate Professor, University of New Haven
- Khaled Sayed, Ph.D.
- Committee Member, Assistant Professor, Electrical & Computer Engineering and Computer Science, University of New Haven
- Muhammad Aminul Islam, Ph.D.
- Committee Member, Assistant Professor, Electrical & Computer Engineering and Computer Science, University of New Haven
- Xenophon Papademetris, Ph.D.
- External Committee Member, Professor of Biomedical Informatics & Data Science, Yale School of Medicine
Clinical collaborators
These two clinicians are coauthors on the benchmark. The 1,200-pair equivalence sample that the automated judge is measured against was adjudicated by three reviewers: a radiologist, a clinician, and a device-research reviewer.
- Dr. Lekshmy Jayan
- Department of Radio-Diagnosis, Mount Zion Medical College Hospital, Adoor, Kerala, India
- Dr. Arun Gopinatha Kurup
- Badr Al Samaa Hospitals, Nizwa, Muscat, Oman
Invited talk
Consistency is not safety in medical vision-language models · AI Research Symposium (HAIDRS) 2026, Hartford AI Day, hosted by Connecticut AI Alliance. Slides.
Glossary
The site uses a few terms of art, and one of them travels under more than one name.
- Paraphrase flip
- The same image and a clinically equivalent rewording of the same question produce a different answer. The pairwise flip rate counts flips per question-paraphrase pair; query-level rates count a question as flipped if any rewording flips it, and run higher.
- Image-invariant
- The answer does not change when the image is removed or swapped. Also called text-reliant or text-driven on this site and in the papers: all three name the same behaviour.
- Text prior
- The mechanism behind image-invariance: the answer the question text alone would produce, learned from how often findings appear in training data.
- Patient-disjoint
- No patient appears in both training and evaluation, so the model cannot pass by recognising a person it has seen.
- AUROC
- Area under the receiver operating characteristic curve: how well a signal ranks one class above another. 0.5 is a coin flip, 1.0 is perfect.
- AUGRC
- Area under the generalized risk-coverage curve, a selective-prediction quality score. It runs 0 to 0.5, and lower is better.
- Predictive entropy
- How close the model's yes/no probability is to a coin flip, from one forward pass. High entropy means low confidence.
- Equivalence audit
- The rubric-based screen that keeps only candidate paraphrases which preserve the clinical question, validated against 1,200 clinician-adjudicated pairs.
Scope of this research
This work measures binary yes/no clinical questions on chest radiographs, across three datasets from three continents, offline and retrospectively. The mechanistic work is on a single model family. The full set of bounds is on the thesis map.
Data and image licences
The three chest X-ray corpora carry their own terms, and this site publishes none of their images.
| Dataset | How to get it | Licence |
|---|---|---|
| MIMIC-CXR | Credentialed PhysioNet access (CITI training plus a signed data use agreement); redistribution of images or reports is prohibited, so the benchmark releases only question/paraphrase text and identifiers. | PhysioNet Credentialed Health Data License 1.5.0 |
| PadChest | Registration with the Medical Imaging Databank of the Valencia Region (BIMCV) is required before download. The PadChest Dataset Research Use Agreement grants research use only and prohibits reproducing or publishing any portion of the dataset without prior written permission, so the benchmark releases only question and paraphrase text. | PadChest Dataset Research Use Agreement (BIMCV; research use only, redistribution prohibited without written permission) |
| VinDr-CXR | Credentialed PhysioNet access (CITI training plus a signed data use agreement); redistribution prohibited. | PhysioNet Credentialed Health Data License 1.5.0 |
| CheXlocalize | Free registration with Stanford AIMI is required before download; released for research use. | Stanford AIMI dataset terms (research use) |
The PSF-Med release contains questions, paraphrases, judge labels, and audit verdicts. It contains no images. The failure gallery shows no radiographs either: its cases are built on MIMIC-CXR studies, whose licence prohibits redistribution, so each case is described in words instead. The only images from these corpora anywhere on this site are the figure panels inside the dissertation PDF. Nothing on this site carries a patient identifier, a study identifier, or a source filename: the data layer is checked for those at build time and the build fails if one appears.
Accessibility
This site targets WCAG 2.2 AA.
What is in place
Every chart has a data table and a text description. No information is carried by color alone: flagged values also carry a hatch pattern and a marker, and series use different shapes. Every control is reachable and operable by keyboard, focus is visible and restored after dialogs, and motion respects a reduced-motion preference. Both themes meet the contrast minimums.
Known gap
The dissertation PDF itself is not tagged for screen readers. Until it is remediated, the abstract, the table of contents, every result, every chart, and every citation are available here as accessible HTML instead.
If something on this site is not usable for you, tell me and I will fix it.
How this site is built
Every statistic on this site comes from a central data layer. None is typed into a page by hand. Each record has to name its metric, denominator, sample size, population, source file, and the limit on its interpretation. The build fails if any of those is missing, if a claim points at a result that does not exist, or if a paper marked as under review exposes a manuscript link.