Table 3.
Classification accuracy values obtained for the two CASE STUDIES (CASE STUDY 1, CASE STUDY 2) by applying METHOD A (time series analysis by 1D Deep Learning strategy as that described in17) and METHOD B (classifiers using basic trajectory features).
| ACCURACY VALUES | Single-cell analysis | Tumor-cell microenvironment analysis (majority voting) | Video-level analysis (majority voting) |
|---|---|---|---|
| CASE STUDY 1 | |||
| METHOD A | 59% (55–62%) | 59% (55–62%) | 59% (50–67%) |
| METHOD B | 64% (60–70%) | 79% (70–87%) | 75% (67–83%) |
| ACCURACY VALUES | Single-cell analysis | Cluster-level analysis (majority voting) | Video-level analysis (majority voting) |
| CASE STUDY 2 | |||
| METHOD A | 55% (54–56%) | 55% (46–63%) | 71% (67–75%) |
| METHOD B | 65% (63– 66%) | 70% (68–72%) | 71% (67–75%) |
The three diverse consensus levels of Tab. 2 have been considered. First column indicates single-cell classification results with no-consensus; second column indicates tumor-cell environment analysis where consensus by majority voting is performed over all the immune cells tracks in the neighbourhood of the same cancer cell (case study 1) or cancer cells tracks belonging to the same cluster (case study 2); third column indicates that majority voting has been performed at the video level by combining the prediction of all the tracks within the same video. Balanced accuracy has been evaluated for all the situations to account for samples unbalance. The accuracy values within the brackets indicate the results obtained in each turn of the two-fold validation procedure. The average accuracy values are also reported.