Tanha Asfika Jaman

AI researcher · medical imaging · reliable machine learning

BSc student in Artificial Intelligence · NUIST, Nanjing, China

I study when medical-imaging models deserve to be trusted — whether their confidence survives a new scanner, whether a benchmark leaks, and whether a result can be reproduced by someone else.

Explore research View publications GitHub Contact
Portrait of Tanha Asfika Jaman wearing a black blazer, standing outdoors in warm afternoon light
Fig. 0 Tanha Asfika Jaman · Nanjing
New Published in IJCA Vol. 187 No. 149 — Does Probability Calibration Survive a Change of OCT Scanner?

§01About

Reliability before accuracy.

Head-and-shoulders portrait of Tanha Asfika Jaman
Affiliation School of Artificial Intelligence
Nanjing University of Information Science and Technology (NUIST)

I am an undergraduate in Artificial Intelligence at the School of Artificial Intelligence, Nanjing University of Information Science and Technology (NUIST), in Nanjing, China.

My research sits where deep learning meets medical imaging, and it is mostly about reliability rather than leaderboard accuracy. As first author, I asked whether the probability calibration of retinal OCT classifiers survives a change of scanner manufacturer — after first auditing the datasets for patient and duplicate-image leakage. The study is published in the International Journal of Computer Applications (2026).

Alongside that work I build audit-style projects on head-CT and brain-MRI benchmarks, a network-science project on influence maximization, and I have co-authored papers on zero-shot learning for remote-sensing scenes and radar signals. I care about leakage-free splits, pre-stated hypotheses, negative results reported honestly, and code that others can run.

  1. Co-authored CLIP-based zero-shot classification of imbalanced remote-sensing scenes — INJIISCOM
  2. Co-authored Zero-shot STFT radar & communication signal classification — Discover Electronics
  3. Learning Courses in AI, data science, responsible AI and cybersecurity (17 completions)
  4. Code Public research repositories on OCT calibration, head-CT audit, brain MRI and graph influence
  5. First author Calibration transfer across OCT scanner manufacturers — published in IJCA

§02Research focus

What I work on.

  • 01

    Calibration & reliability

    Whether a model's confidence means what it says — temperature scaling, ECE and risk at a referral threshold, especially under domain shift.

    Evidence: IJCA 2026oct-calibration-transfer

  • 02

    Medical image analysis

    Retinal OCT, non-contrast head CT and brain MRI classification with transfer-learned CNN backbones.

    Evidence: IJCA 2026accuracy-without-anatomyBrainTumorFusion

  • 03

    Dataset audits & leakage

    Patient-disjoint and near-duplicate-aware splits, shortcut detection, and checking what a benchmark actually measures.

    Evidence: IJCA 2026accuracy-without-anatomy

  • 04

    Reproducible ML

    Pinned environments, seeded runs, tests, and scripts that verify every number in a manuscript against saved results.

    Evidence: oct-calibration-transferaccuracy-without-anatomy

  • 05

    Zero-shot & vision–language

    Co-authored work on CLIP-based zero-shot scene classification and zero-shot radar/communication signal recognition.

    Evidence: INJIISCOM 2025Discover Electronics 2025

  • 06

    Network science

    Community-aware influence maximization on social graphs with Louvain/Leiden and multi-centrality fusion.

    Evidence: hybrid-caim-facebook

§03Publications

Published work.

Published · peer-reviewed First author Journal article

Does Probability Calibration Survive a Change of OCT Scanner? A Leakage-Free Study of Calibration Transfer Across Manufacturers

Tanha Asfika Jaman, Akter Mst Lovely, Iftee Shekh Iftesham, Mostafa Farzana

International Journal of Computer Applications · Vol. 187 · No. 149 · pp. 37–44 · October 2026 · ISSN 0975-8887

Read on IJCA PDF DOI 10.5120/ijca8fee57096b16 Code

Research question

Do the predicted probabilities of retinal OCT classifiers stay calibrated when the scanner manufacturer changes?

Study

Three-class OCT classifiers (AMD, DME, normal) were trained on patient-disjoint Heidelberg Spectralis B-scans from the Kermany dataset and evaluated on an unseen Optovue Avanti RTVue XR cohort (OCTDL), across eleven runs, two backbones and two resizing policies. Before any modelling, the study audited both datasets for leakage and built splits that keep the same patient and near-duplicate images together.

Method

  • Leakage audit of both public datasets: shared patient identifiers across Kermany class folders, and pixel-identical OCTDL images filed under different patient IDs.
  • Splits built from connected components of “same patient” or “near-duplicate” relations.
  • Source-only temperature scaling compared with physics-parameterised OCT acquisition degradation, a matched-severity generic corruption family, and input-conditioned temperature.

Takeaway

Resizing policy mattered more than any calibration method tested, a single source-fitted temperature was the most dependable fix, and the error promised at a source-calibrated referral threshold more than doubled on the new scanner.

Optical coherence tomographyProbability calibrationDomain shiftData leakageTemperature scalingRetinal imagingReproducibility

Findings reported in the published abstract

Target balanced accuracy

Letterbox resize0.8821
Square resize0.9304

Higher is better

Uncalibrated target ECE

Letterbox resize0.0316
Square resize0.0181

Lower is better

10/11 runs

Source-only temperature scaling lowered target ECE

6/11 runs

…with a 95% bootstrap interval excluding zero

5/11 runs

Physics-based OCT degradation beat a generic corruption family

10/11 runs

Input-conditioned temperature lost to one global temperature

A referral threshold promising 0.5% error on the source scanner gave 1.19%–1.59% error on the new scanner.

Co-authored articles

2025

Experimental Evaluation of CLIP-Based Zero-Shot Classification of Imbalanced Remote Sensing Scenes: Addressing Quantity Disparities in Data

Tanvir Ahmed, Tanha Asfika Jaman, Shekh Ifteesham Iftee, Tanjoy Mahmud, Ekra MD Emadur Rahman, Hossain MD Maruf

International Journal of Informatics, Information System and Computer Engineering (INJIISCOM) · Vol. 6 · No. 1 · pp. 130–143 · 2025

A zero-shot remote-sensing scene classification framework built on CLIP and Vision Transformer backbones, aimed at linking visual and semantic features when classes have very different numbers of images; evaluated on WHU-RS19.

Co-author (second author) Article PDF

Author names appear exactly as printed by each publisher. Profiles: ORCID · ResearchGate · Google Scholar.

§04Projects

Open research code.

Research code · published study

Calibration transfer across OCT scanners (opens GitHub)

The research artifact behind the IJCA paper: a leakage-free protocol, dataset audits, two degradation families and five calibration conditions for retinal OCT classifiers moving from Heidelberg Spectralis to Optovue scanners.

  • Splits grouped by patient and near-duplicate images (NCC ≥ 0.98)
  • ResNet-18 and ConvNeXt-Tiny as measurement instruments, configurations analysed separately
  • Analysis-only path reproduces the tables without a GPU; 66 pytest tests

PythonPyTorchtimmpytest

First-author studyMIT (code)

Benchmark audit · head CT

Accuracy Without Anatomy (opens GitHub)

A redundancy and shortcut audit of a public non-contrast head-CT stroke benchmark: perceptual near-duplicate detection, redundancy-aware splits, brain-only vs brain-deleted inputs, a permuted-label control and Grad-CAM scored against expert lesion masks.

  • Tests whether classifiers rely on signal outside the intracranial compartment
  • Random vs redundancy-aware split policies over three seeds
  • Script that checks every manuscript number against saved results

PythonPyTorchResNet-18Grad-CAM

Corresponding author of the accompanying manuscriptMIT (code)

Medical imaging · brain MRI

BrainTumorFusion (opens GitHub)

Research software for four-class brain MRI classification (glioma, meningioma, pituitary, no tumour) with a reproducible evaluation path: hash-based duplicate-safe splits, validation-only checkpoint selection, one final test run, confusion matrices and Grad-CAM overlays.

  • Configurable ResNet-18/50 encoders with an image-branch fusion block
  • Streamlit demo and Tkinter viewer for local inspection
  • Performance claims deliberately withheld until the clean pipeline is re-run

PythonPyTorchtorchvisionStreamlit

AuthorMIT (code)

Network science · influence

Hybrid CAIM (opens GitHub)

Community-aware influence maximization on the SNAP Facebook ego-network: Louvain and Leiden communities, a five-measure centrality fusion score, one seed per community, and spread measured with from-scratch Independent Cascade and Linear Threshold models.

  • Compared against random, degree and PageRank seeding
  • Trades a little spread for guaranteed coverage of every community
  • Accompanied by a technical report (Dept. of AI, NUIST, 2026)

PythonNetworkXleidenalgigraph

AuthorMIT (code)

§05Certificates & learning

Continuous learning.

17 unique course and learning-path completions from Huawei ICT Academy, HP LIFE, LinkedIn Learning, Cisco Networking Academy and Google on Coursera. These are course completions — not degrees and not proctored professional certification exams.

Duplicate copies of the same course are shown once (20 files → 17 records). Credential IDs are withheld on previews; issuer verification links are included where the original certificate provides one.

§06Skills

Tools, tied to evidence.

Machine learning & deep learning

  • PyTorch
  • torchvision
  • timm
  • Transfer learning (ResNet, ConvNeXt)
  • Mixed-precision training
  • AdamW / one-cycle & cosine schedules

Used in oct-calibration-transfer · accuracy-without-anatomy · BrainTumorFusion

Reliability & evaluation

  • Probability calibration (ECE, temperature scaling)
  • Domain-shift evaluation
  • Bootstrap confidence intervals
  • Leakage-free & group-aware splits
  • Near-duplicate detection
  • Grad-CAM attribution

Used in IJCA 2026 · accuracy-without-anatomy

Medical imaging

  • Retinal OCT
  • Non-contrast head CT
  • Brain MRI
  • Dataset licensing & provenance checks

Used in IJCA 2026 · accuracy-without-anatomy · BrainTumorFusion

Data science & graphs

  • Python
  • NetworkX
  • Community detection (Louvain, Leiden)
  • Centrality analysis
  • Diffusion models (IC, LT)
  • Matplotlib & Seaborn

Used in hybrid-caim-facebook · Cisco Data Science Essentials with Python · Google Foundations: Data

Reproducible research practice

  • Git & GitHub
  • pytest
  • Pinned environments (conda, requirements)
  • Seeded multi-run experiments
  • Number-verification scripts
  • CITATION.cff & licence hygiene

Used in oct-calibration-transfer · accuracy-without-anatomy

Responsible & applied AI

  • Generative AI fundamentals
  • AI ethics
  • Critical evaluation of AI output
  • Cybersecurity awareness

Used in LinkedIn Learning · HP LIFE · Cisco Networking Academy courses

Languages: English · Bangla (Bengali)

§07Contact

Let’s talk research.

Open to conversations with supervisors, research groups and collaborators working on reliable medical AI, calibration, dataset auditing and reproducible evaluation.