SAM for medical imaging

This is Segment Anything (SAM) on DICOM: Meta’s promptable model, a small adapter that turns a DICOM slice into the RGB array SAM expects, and a lung-mask example. It is not a generic medical-image-segmentation explainer — that is medical image segmentation. It is not SAMJ inside Fiji — that plugin is on the ImageJ page.

What SAM is

Meta’s Segment Anything Model is a promptable segmenter. You give it an image plus a point, a box, text, or a coarse mask. It returns a mask. Three parts:

• Image encoder — a Vision Transformer (ViT) on a high-resolution input
• Prompt encoder — sparse prompts (points, boxes, text) and dense prompts (masks)
• Mask decoder — image embedding + prompt embeddings → a mask

The point of the design is zero-shot transfer: new image types without retraining the whole stack. Medical CT is still a new image type. Upstream SAM eats PNG and JPEG, not DICOM and not Hounsfield units.

DICOM adapter

The fork is amine0110/SAM-Medical-Imaging. The missing piece is a reader that turns a DICOM into 8-bit BGR. Min–max normalize the pixel array, scale to 0–255, then COLOR_GRAY2BGR. (The first version of this snippet had broken names: HOUNDSFILD_MIN, HOUNS_MIN, normalized. Use this listing.)

def prepare_dicoms(dcm_file, show=False):
    pixel = pydicom.dcmread(dcm_file).pixel_array.astype(np.float32)
    hu_min = float(np.min(pixel))
    hu_max = float(np.max(pixel))
    hu_range = max(hu_max - hu_min, 1.0)
    normalized = (pixel - hu_min) / hu_range
    uint8_image = np.uint8(normalized * 255)
    bgr = cv2.cvtColor(uint8_image, cv2.COLOR_GRAY2BGR)
    if show:
        cv2_imshow(bgr)  # Google Colab
    return bgr

That is a window, not a clinical HU preset. For lung you will usually window first (e.g. centre −600, width 1500) and then scale. SAM still wants three channels.

Output

Prompt the model on the prepared slice (a click in the lung, or a box). The mask below is a DICOM chest slice with the lung filled.

Code: github.com/amine0110/SAM-Medical-Imaging. Fiji users who want SAM as a click tool should use SAMJ on the ImageJ page, not this Python fork.

PYCAD builds the imaging side of this when the mask has to live in a clinic app. Case studies.