Fig 1.
Study population characteristics.
Table 1.
GLCM textural features used.
Fig 2.
Sequential domain reduction transformer on the “max depth” hyperparameter.
Table 2.
Patient demographics and sample sizes.
Fig 3.
Post-contrast MRIs from DCE for (A) a pCR patient and (B) a non-PCR patient at the pre-treatment, early treatment, and mid-treatment time points. For (A), the 54 year old Caucasian patient had invasive ductal carcinoma with HR- and HER2-. For (B), the 51 year old Caucasian patient had invasive ductal carcinoma with HR+ and HER2-. The yellow arrow indicates the approximate location of the tumor. MRI images were taken approx. 2.5 minutes after contrast injection with a phase duration of 80sec < phase duration < 100sec. pCR: complete pathological response, DCE: dynamic contrast enhanced.
Fig 4.
DWIs for (A) a pCR patient and (B) a non-PCR patient at the pre-treatment, early treatment, and mid-treatment time points. For (A), the 54 year old Caucasian patient had invasive ductal carcinoma with HR- and HER2-. For (B), the 51 year old Caucasian patient had invasive ductal carcinoma with HR+ and negative HER2-. The yellow arrow indicates the approximate location of the tumor. DWI acquisitions were done using a b-value of 800 s/mm2. DWI: diffusion-weighted imaging).
Fig 5.
Top 10 most important texture features out of 372 features by F-score.
GLCM features are shown with pixel distance value (1, 3 or 5 pixels), direction (North/South, East/West, North East/South West, or North West/South East), and MRI imaging type (DCE: dynamic contrast enhanced or DWI: diffusion-weighted imaging). GLCM: gray-level co-occurrence matrix.
Fig 6.
Mean ADC values for all non-pCR and pCR patients at the pre-treatment, early treatment, and mid-treatment time points.
Error bars are standard deviations. pCR: complete pathological response.
Table 3.
Model performances for models trained on ADC, DCE(GLCM), DWI(GLCM), their combinations, all non-imaging data, and the combination of all MRI and all non-MRI data with 5-fold cross-validation.
Table 4.
Model performances for models trained on all features restricted to tp0, tp1, tp2, tp0 + tp1 and all timepoints using 5-fold cross-validation.