fidoptMMOTU CEUS, ovarian tumours in contrast-enhanced ultrasound
A contrast-enhanced ultrasound image of an ovarian tumour, 2D and in colour, with the tumour outlined by a gynaecology expert and checked by a second (MMOTU's OTU_CEUS subset, Beijing Shijitan Hospital).
10 test casespelvisimage US2D, distances in pxresearch only
Read before comparing (10)
- Images. JPEGs 330 to 888 pixels wide and 218 to 657 high, cropped by hand to remove the patients' details, in colour as stored; no pixel size is given, so distances are in pixels. The folder is named OTU_3d, yet the images are 2D CEUS frames. The paper names a Mindray Resona8, but some images carry a GE LOGIQ E9 banner (ceus_100) and others Mindray's (ceus_001).
- Hints. Arrows the sonographers drew around the lesion (ceus_001) and text naming it ('OV-CY' on ceus_059) are burnt into many images; the truth follows them on the images looked at.
- Truth. Mostly one tumour an image, of seven types (this subset has no normal ovary): test case ceus_021 and the held-out ceus_159 and ceus_160 outline two cysts, one in each ovary, test case ceus_115 outlines three locules, and the held-out ceus_134 shows a second cyst that is not outlined. The same frame is filed under two types, 104 to 107 as simple cysts and 164 to 167 as chocolate cysts. Only the binary mask is scored, not the type.
- Groups. In file-number order the 170 images fall into 37 runs of one tumour type, and most runs hold frames of one or a few lesions. Image correlation (normalised, on 64-pixel thumbnails) gives the reviewer's nine known same-lesion pairs 0.60 to 0.89; of the 63 pairs across runs at 0.60 or more, the 21 joining 98-110 to 151-170 are the same frames (0.72 to 1.00) and the others show other lesions by eye. The runs joined across that block make the 31 groups; a run that holds several lesions stays one group.
- Copies. Images of one lesion are often frames seconds apart: 15 pairs inside the provider's test split are near copies (correlation 0.95 or more, such as 159 and 160, 56 and 59, 42 and 48), and 151 to 170 repeat 98 to 110, with measurement marks added on 158. One image a group keeps none twice.
- Patients. MMOTU holds 294 patients; the paper's counts give 247 to OTU_2d, its 1469 B-mode images, and 47 to this subset, which adds up, but that rests on the same paper whose patient-disjoint split the images contradict, so the subsets may share patients. File names carry no patient: a group may hold several patients, or a patient two groups.
- Overlap. No other dataset of the corpus holds ovarian ultrasound; SonoCorpus's MD5 manifest finds none of these images in its other 52 datasets, OTU_2d included.
- Outside numbers. The masks are public since 2022, and a model trained on OTU_CEUS's own training split has seen the test lesions, its images being frames of the same lesions. SonoBase trained on OTU_2d and was evaluated on this subset, not trained on it.
- Start. On the 10 test images the tumour covers 15% of an image on average: the whole image scores a Dice of 0.26, Otsu's dark pixels 0.27 and its bright ones 0.09. TotalSegmentator has no ultrasound task.
- Tolerance. 2 pixels, as KvasirSEG on images of similar width: the truth grown or thinned by one pixel keeps a Dice of 0.99 and a surface Dice of 1.0 at 2 pixels.
No method has a result on every case of this dataset yet.
Case by case
Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- The provider's split is not used: its training and test images alternate inside runs of one tumour type and show the same lesions (train_18 and test_20, train_23 and test_21), and images 151 to 170 repeat frames of 98 to 110. The 170 images form 31 lesion groups (source.json's group), one image a group, the lowest-numbered: 10 test groups drawn with seed 2026, then 10 development groups; every other image is in draw.held_out.
- Truth
- contours by experts the provider's binary masks (<n>_binary.PNG), each image outlined pixel by pixel by one of 27 experts of the hospital's department of gynaecology and obstetrics and checked by another, with the pathology report at hand (Zhao et al., arXiv 2207.06799, section 3.1)
- Source
- github.com/cv516Buaa/MMOTU_DS2Net/tree/be8455e70d2f73bdc9143236eacce4b5b38b604c
drive.google.com/drive/folders/1c5n0fVKrM9-SZE1kacTXPt1pt844iAs1
arxiv.org/abs/2207.06799
doi.org/10.1016/j.patcog.2025.112311
doi.org/10.5281/zenodo.22770825 - Licence
- none stated for the data (the repository's LICENSE, Apache 2.0, is MMSegmentation's, for the code)
10 cases for tuning, never tested
ceus_098, ceus_108, ceus_101, ceus_001, ceus_061, ceus_125, ceus_066, ceus_081, ceus_007, ceus_136