Marcin Stepien

Software Engineer, posts about tech, complexity, art.

[Marcin Stepien @marcinstepien]

Copernicus Festival 19 - 24 May 2026 is about #complexity
Among speakers Andrzej Udalski, David Krakauer @sfiscience, Hiranya Peiris @uclmaps, @johnhawks, @anilkseth, Jakub Zakrzewski, @PaluchLab, Leszek Kaczmarek @NenckiInstitute, @piotrsankowski
#CopernicusFestival free & open

[Marcin Stepien @marcinstepien]

Whether you are an emergentist, a dualist, or a strong AI believer, we would like to welcome you to another event of #Complexity Explorers Krakow, set for 28 May. This time we will #debate on #computationalist s' views together with Radek Kycia, @duda_jarek. Attendance is free.

[Marcin Stepien @marcinstepien]

4. And the last but crucial one: #privacy.
At just ~4GB for inference, we are finally reaching a point where privacy-sensitive medical images can be processed locally. No data needs to leave the user's device to get a high-quality initial screening. #MedSigLIP

[Marcin Stepien @marcinstepien]

Have been working with the MedGemma lately, specifically on MedSigLIP for Medical Image Interpretation, a dermatology use case of identifying melanoma from mobile-captured photos.
Here's a thread for my findings
#MedTech #HAI-DEF #HealthCare #Dermatology

[Marcin Stepien @marcinstepien]

- Preserve the aspect ratio of an object on image. Since asymmetry is a primary indicator of melanoma, distortion can destroy the model's ability to "see" the pathology.

[Marcin Stepien @marcinstepien]

3. Image preprocessing.
The model handles up to 448x448 px. To maximize accuracy:

[Marcin Stepien @marcinstepien]

- Use object detection to center the lesion and crop the image before feeding it to the model.

[Marcin Stepien @marcinstepien]

- The Bias: Accuracy varies significantly across different skin tones. Research indicates a historical performance gap in derm-AI; for example, models trained on standard datasets (like HAM10000, which is ~95% Caucasian-centric) can see a 10-15% drop in diagnostic accuracy

[Marcin Stepien @marcinstepien]

2. Hospital vs. real world data.
MedSigLIP was primarily trained on hospital-grade, contrast-enhanced, and DICOM-derived images.

[Marcin Stepien @marcinstepien]

- The Issue: Mobile photos lack the controlled lighting and focus of clinical imaging.

[Marcin Stepien @marcinstepien]

- Select classes.
Not all conditions have the same accuracy in classification. Focus your classes on high-accuracy clusters rather than a broad net.

[Marcin Stepien @marcinstepien]

1. The zero-shot challenge.
MedSigLIP aligns images and text in a shared embedding space, but out-of-the-box results can be prone to false positives. To mitigate this:

[Marcin Stepien @marcinstepien]

because it mathematically maps those specific words to the visual features of redness and scaling it learned during pre-training. You may want to experiment with medical triage to rule out false positives (I tried but eventually get better result without)

[Marcin Stepien @marcinstepien]

- Add morphological descriptors detailing shape, color, border, arrangement, and texture.
E.g, when MedSigLIP reads the word "Psoriasis", its text encoder activates much stronger when it also reads "erythematous plaques with silvery-white scale",

[Marcin Stepien @marcinstepien]

- Vocabulary matters.
Use formal medical terminology the model was trained on (like in PubMed). If you use "common" language, standard models like SigLIP 2 actually tend to outperform the specialized MedSigLIP.

[Marcin Stepien @marcinstepien]

- Prompt engineering is mandatory.
Don't just provide a label. Attach short descriptions of the characteristics of an image. Add a contextual prefix, e.g. instead of  "Photo of .." put "Dermoscopy image revealing ..."

[Marcin Stepien @marcinstepien]

In the mood for creativity. Where many builders are right now.

[Uncle Bob Martin @unclebobmartin]

I do not feel like I'm not programming. In fact, I feel like I am programming more intensely than ever. I'm just not coding.

[Marcin Stepien @marcinstepien]

My build of Dermatolog AI Scan application with MedSigLIP model https://t.co/qkuvAWKOki

[Marcin Stepien @marcinstepien]

And so moved through the rewarding pursuits of philosophy and artistic painting before landing on the raw reality of technical algorithms and a concrete application at the later section of this text at https://t.co/Ehaa2cyryM

[Marcin Stepien @marcinstepien]

In recent years I have returned to painting on canvas. It improves seeing. Extends focus beyond daily analytical work in IT. There is also an emotional aspect, some truth to how Winston Churchill famously used painting as a therapeutic escape from his "Black Dog".