sw/
← projects
VIBOT MSc·2017

Breast tomosynthesis density classification

Classifying breast density into the four BI-RADS classes from digital breast tomosynthesis, using Local Binary Patterns and k-NN. 93.75% test accuracy and first place in the VIBOT Medical Image Analysis contest.

Problem

Dense breast tissue raises the risk of breast cancer and makes it harder to spot, so radiologists grade it on the four-class BI-RADS scale, from almost fatty (under 25% glandular tissue) to extremely dense (over 75%). The grading is subjective, and different radiologists can rate the same image differently. Digital breast tomosynthesis (DBT) gives a 3D stack of slices rather than a single mammogram, which helps detection but makes automatic tissue classification harder because of low vertical resolution and out-of-plane artefacts.

What I did

With Daniel González-Adell and Ziyang Hong, for the Medical Image Analysis course in the VIBOT master at the Universitat de Girona, I built a pipeline that grades a DBT volume into BI-RADS automatically. We first tried texture features from grey-level co-occurrence matrices, regions of interest and SIFT descriptors, but none of them separated the classes well.

What worked was Local Binary Patterns, an edge and texture descriptor that suits the granular look of dense tissue. Each slice is cropped to the breast and normalised by its maximum intensity, then LBP is computed with a radius of 20 pixels and 9 neighbours. The LBP image is binarised with Otsu’s method, the breast is masked out, and a morphological opening with a disk of radius 6 removes the skin line. The descriptor is simply the ratio of LBP edge pixels to all pixels inside the breast, and only the densest slice of each volume is used.

I then compared classifiers in MATLAB: k-nearest neighbours and a multiclass SVM built from one-versus-one binary SVMs, evaluated with confusion matrices, sensitivity, specificity, precision and the Dice coefficient for each class.

Result

Trained on 16 DBT volumes and tested on 16 others, k-NN reached 93.75% accuracy, getting 15 of the 16 test volumes right, against 63% for the multiclass SVM. Being a single ratio per volume, the method is also fast. It placed first in the VIBOT Medical Image Analysis contest on breast density estimation, held in Girona on 25 May 2017. The obvious next steps were better pre-processing, more training data than 16 volumes, and LBP-TOP to follow dense tissue across slices.

diagrams

scroll sideways on small screens

Computing one Local Binary Pattern codeA pixel of value 70 and its eight neighbours 70, 78, 83, 80, 62, 47, 51 and 65. Subtracting the centre gives 0, 8, 13, 10, minus 8, minus 23, minus 19 and minus 5. Thresholding at zero gives 1, 1, 1, 1, 0, 0, 0, 0. Weighting by powers of two and summing gives 15.LBPP,R = Σp=0..P−1 s(gp − gc) · 2p s(x) = 1 if x ≥ 0, else 07070788380624751651 · sample0+8+13+10-8-23-19-52 · difference from centre1201211221230240250260273 · threshold at 01·1 + 1·2 + 1·4 + 1·8 + 0·16 + 0·32 + 0·64 + 0·128 = 154 · weight by powers of two and sum
diagramHow one Local Binary Pattern code is computed: compare each of the 8 neighbours with the centre, keep 1 where it is at least as bright, and read the bits as a binary number.Source: the 2017 course report
k-NN confusion matrixRows are the true BI-RADS class, columns the automatic classification. All four BI-RADS I, III and IV volumes are correct; three of four BI-RADS II volumes are correct and one is classified as BI-RADS III. 15 of 16 correct.Automatic classificationTruthBI-RADS IBI-RADS IIBI-RADS IIIBI-RADS IVBI-RADS IBI-RADS IIBI-RADS IIIBI-RADS IV400003100040000415 of 16 test volumes correct · accuracy 93.75%
resultk-NN confusion matrix on the 16 test volumes. The diagonal is correct; the only error is one BI-RADS II volume read as BI-RADS III.Data from the 2017 course report
k-NN metrics per BI-RADS classPrecision, sensitivity, specificity and Dice coefficient are all 1.00 for BI-RADS I and IV. BI-RADS II: precision 1.00, sensitivity 0.75, specificity 1.00, Dice 0.86. BI-RADS III: precision 0.80, sensitivity 1.00, specificity 0.92, Dice 0.89.0.000.250.500.751.00BI-RADS I0.750.86BI-RADS II0.800.920.89BI-RADS IIIBI-RADS IVPrecisionSensitivitySpecificityDice
resultk-NN precision, sensitivity, specificity and Dice coefficient per class. Unlabelled bars are 1.00; the single II-as-III error is what pulls down II's sensitivity and III's precision.Data from the 2017 course report

figures

click to enlarge

fig. 01First prize, Breast Tomosynthesis density classification challenge, VIBOT Medical Image Analysis module. Girona, 24 May 2017.Universitat de Girona · ViCOROB
fig. 02Original slice and LBP image for BI-RADS I, II, III and IV, top to bottom. The denser the breast, the more LBP texture appears.Course report, 2017
esc