Image credits

The homepage film and model images are crops of figures from each package’s tagged documentation, listed in the source manifest. The crops are placed on black, slightly desaturated and animated with slow camera movement and dissolves. Their predictions are unedited.

The radiomics image and the Segmentation, Detection and Quantitative imaging card videos were computed for this site, and each has a run record. In those videos the CT is windowed, cropped and resized, pixels outside the body are set to black, and model outputs are unedited.

Abdominal CT

Abdominal CT documentation figures other than TotalSegmentator’s use AMOS22, by Yuanfeng Ji et al. (2022). The Zenodo record lists CC BY 4.0 and the archive includes CC BY-SA 4.0 terms, so these figures and their crops use CC BY-SA 4.0.

TotalSegmentator

The TotalSegmentator surface rendering and axial CT panel use autoPET-III-Lite, a subset of the autoPET III challenge data, under CC BY-NC 4.0. Both panels and their crops keep those terms.

Brain MRI

The brain tumor MRI and its SegResNet segmentation come from the Medical Segmentation Decathlon brain tumor task (BraTS), by Antonelli et al. (2022), under CC BY-SA 4.0. mlx-generate-ctmr generated the synthetic T1 image with the MR-Brain configuration of NVIDIA NV-Generate. The SynthSeg parcellations are of that image.

Chest CT and radiograph

Nodule detection, MedSAM2 and the Detection card use LUNA16, by Bram van Ginneken and Colin Jacobs, under CC BY 4.0. Setio, A. A. A., et al. (2017). Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the LUNA16 challenge. Medical Image Analysis 42, 1–13.

LUNA16 is drawn from LIDC-IDRI, under CC BY 3.0: Armato III, S. G., et al. (2015). Data From LIDC-IDRI [Data set]. The Cancer Imaging Archive. doi:10.7937/K9/TCIA.2015.LO9QL9SX.

The FleXray rib segmentation uses Mikael Häggström’s normal PA chest radiograph, CC0 1.0.

FDG PET/CT

The FDG PET/CT panel uses defaced CT from Gatidis, S., Kuestner, T. (2022). A whole-body FDG-PET/CT dataset with manually annotated tumor lesions (FDG-PET-CT-Lesions) (Version 2) [Data set]. The Cancer Imaging Archive. doi:10.7937/gkr0-xv29, under CC BY 4.0. The package simulates the PET overlay and uptake focus; they are not measured PET.

Radiomics image

The film’s radiomics panel is a local entropy map that mlxradiomics 0.1.0.dev5 computed through MedMLX on the Apple Silicon GPU. Its input is an AMOS22 axial CT slice (amos_0008, slice 80) and the SwinUNETR liver prediction. The ROI is the largest of the prediction’s three connected components; the other two were excluded before extraction. The run record lists the selection, input hashes, extraction settings and display mapping.

The map uses 10-HU bins and an 11 × 11 in-plane neighborhood. Color shows entropy in bits, scaled to the ROI’s 2nd–98th percentiles. The map is not smoothed. It is a single-slice demonstration with no clinical interpretation.

Segmentation card

The Segmentation card scrolls through 35 axial slices of case HCC_063 from the HCC-TACE-Seg collection: Moawad, A. W., Fuentes, D., Morshid, A., Khalaf, A. M., Elmohr, M. M., Abusaif, A., Hazle, J. D., Kaseb, A. O., Hassan, M., Mahvash, A., Szklaruk, J., Qayyom, A., & Elsayes, K. (2021). Multimodality annotated HCC cases with and without advanced imaging segmentation [Data set]. The Cancer Imaging Archive. doi:10.7937/TCIA.5FNA-0924, under CC BY 4.0.

The organ labels are SwinUNETR predictions from swin-unetr-mlx 0.2.5. The collection’s reference segmentations are not shown. The case has hepatocellular carcinoma. The labels have not been reviewed and are not a clinical result. The run record lists the series, slice range, settings and hashes.

Detection card

The Detection card scrolls through one LUNA16 subset0 chest CT (series 1.3.6.1.4.1.14519.5.2.1.6279.6001.752756872840730509471096155114) with the top box that lung-nodule-mlx 0.2.5 predicted (score 0.999, center 4.7 mm from the annotated nodule center). The model’s only other box (score 0.03) lies outside the shown slices. The series is on the bundle’s fold-0 validation list, which is held out from training. It was chosen from three large-nodule subset0 cases as the one whose top box was closest to its annotation. The LUNA16 annotation is not drawn. The run record lists the detections, slices and hashes.

Quantitative imaging card

The Quantitative imaging card uses the same AMOS22 slice and SwinUNETR prediction as the radiomics image and adds the kidneys and pancreas. mlxradiomics 0.1.0.dev8 computed one entropy map per organ on the Apple Silicon GPU, from the largest connected part of each predicted organ, with the same settings. Its liver map is identical to the 0.1.0.dev5 map. All frames share one color scale (1.5–2.9 bits). The run record lists each organ’s selection, maps and hashes.

Generation card

The Generation card shows synthetic images only: three CT volumes and three brain MR volumes that mlx-generate-ctmr 0.5.2.post2 generated from noise with NVIDIA NV-Generate weights, under the model terms in its weights repository. No patient image is used. The run record lists the settings, seeds, slice ranges and hashes.

Vision-language card

The Vision-language card shows one LUNA16 subset0 chest CT (series 1.3.6.1.4.1.14519.5.2.1.6279.6001.511347030803753100045216493273), under the LUNA16 and LIDC-IDRI terms above, with an excerpt of the structured report that mlx-reason-ct 0.3.6 wrote for it. The heat shows the decoder attention of each token as it is written, averaged over the full-attention layers; it shows where the model weighted the scan, not why it wrote a word. The report and attention have not been reviewed and are not a clinical result. The run record lists the report, per-token attention peaks, the annotated nodule's position and hashes.

Source images, model weights and derived media retain their respective terms. The website’s Apache-2.0 code license does not relicense them. Return to MedMLX.