fidopt Interactive view
Breast tumours on ultrasound, 2D segmentation

BUSI, breast ultrasound lesions outlined by hand

A 2D B-mode breast ultrasound image of a woman's benign or malignant lesion, outlined freehand and reviewed by radiologists: one of BUSI's 780 images (Baheya Hospital, Cairo, Al-Dhabyani et al. 2020), each group of copies counted once.

30 test casesbreastimage US2D, distances in pxresearch only
Look at the images
Read before comparing (9)
  • Images. 780 PNGs of 600 women aged 25 to 75, from GE LOGIQ E9 and E9 Agile scanners, cropped by the provider to different sizes, about 500 x 500 pixels, stored with three channels (equal where the image is grey) and kept so. No pixel size is given, so distances are in pixels.
  • Copies. Pawłowska, Karwat and Żołek (2023) group the 780 images into 587 views: 153 groups of two to seven copies or near copies, and they object to 44 groups of axilla views, 6 with a biopsy needle and 8 filed under two classes. The pool takes one image a group, its lowest number, and leaves out the objected groups and the normal images, which have no lesion.
  • Truth. Copies of one view carry masks that disagree in shape and place (Pawłowska et al.): an outline is good to a few pixels, not one. 15 images of the pool have two or three masks, joined into one label (benign_58 among the development cases).
  • Women. No patient is named. 587 groups for 600 women suggest about one a woman, but two lesions or two unlike views of one woman could still fall on two sides.
  • Marks. Many images carry text, calipers, Doppler colour or a pictogram burnt in, kept as delivered.
  • Seen. BUSI has been public since 2019 and is in many training sets (SAMUS's US30K, SA-Med2D-20M, BiomedParse). SonoBase trained on SonoCorpus's BUSI training split, which holds 20 of the 30 test images: a model trained on BUSI has probably seen the test cases.
  • Tolerance and start. 2 pixels: on the test images the truth grown or thinned by one pixel keeps a Dice of 0.98 and a surface Dice of 0.99 at 1 pixel, 1.0 at 2. The whole image scores a Dice of 0.16 and its dark pixels (under Otsu's threshold) 0.21, with a surface Dice of 0.03 at 2 pixels; a lesion covers 10% of its image on average (1 to 36%). The test cases hold 18 benign and 12 malignant lesions.
  • Source. The provider's page at Cairo University now asks every visitor to pass a reCAPTCHA. prepare.py reads the original archive from the Wayback Machine's copy of 31 May 2023, every member pinned by the archive's central directory; its members match the MD5s SonoCorpus recorded.
  • Overlap. No other dataset of the corpus holds breast ultrasound. SonoCorpus's MD5 manifest finds one byte copy inside BUSI (benign 433 and malignant 145) and none in its 52 other sets, BUS-BRA, BUSI-WHU and BrEaST among them; it misses 152 of Pawłowska's 153 groups, re-encoded copies, so a re-encoded copy in another set would be missed too.

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
30 test images drawn with seed 2026 and 30 development images, the last of the same seeded order, among the 457 left once each group of copies listed by Pawłowska et al. (2023) counts once and its axilla, needle and two-class groups and the normal images are left out. No patient is named: one image a group stands for one woman.
Truth
contours by experts the lesion outlined freehand in MATLAB, image by image, by people the paper does not name, the annotations reviewed and corrected by radiologists of Baheya Hospital (Al-Dhabyani et al. 2020); an image with two or three masks has them joined
Licence
none stated (the Data in Brief article is CC BY 4.0)
30 cases for tuning, never tested

benign_120, benign_334, malignant_166, benign_241, malignant_133, benign_75, malignant_31, benign_436, benign_312, benign_51, benign_37, malignant_137, benign_1, benign_80, benign_34, benign_361, benign_408, benign_434, malignant_191, benign_58, benign_66, benign_175, malignant_56, benign_12, benign_317, benign_349, malignant_67, malignant_23, malignant_152, malignant_144