Hi MedGemma team,
We are fine-tuning MedGemma 1.5 4B (IT) with LoRA for Turkish radiology report generation using the CT-RATE dataset.
We found the RGB windowing scheme from the fine-tuning notebook:
Red: -1024 to 1024 HU (wide window)
Green: 135 to 215 HU (soft tissue)
Blue: 0 to 80 HU (brain)
Our question:
Should we use these exact same windows when fine-tuning on CT-RATE, or could modifying them improve performance?
Specifically, the blue channel uses a brain window (0–80 HU), but CT-RATE is a chest CT dataset. Would the pretrained SigLIP encoder still expect this specific channel mapping, or is it safe to replace the brain window with something more relevant to chest CT (e.g., a lung window)?
In short: does the model's pretrained representation depend on these exact RGB channel definitions, meaning we must keep them as-is during fine-tuning?
Thanks!
Hi MedGemma team,
We are fine-tuning MedGemma 1.5 4B (IT) with LoRA for Turkish radiology report generation using the CT-RATE dataset.
We found the RGB windowing scheme from the fine-tuning notebook:
Red: -1024 to 1024 HU (wide window)
Green: 135 to 215 HU (soft tissue)
Blue: 0 to 80 HU (brain)
Our question:
Should we use these exact same windows when fine-tuning on CT-RATE, or could modifying them improve performance?
Specifically, the blue channel uses a brain window (0–80 HU), but CT-RATE is a chest CT dataset. Would the pretrained SigLIP encoder still expect this specific channel mapping, or is it safe to replace the brain window with something more relevant to chest CT (e.g., a lung window)?
In short: does the model's pretrained representation depend on these exact RGB channel definitions, meaning we must keep them as-is during fine-tuning?
Thanks!