Comparison with the Works which Classifies into BIRADS Categories

Observing the literature, only the works of Bovis and Singh [17] and Petroudi et al. [137] have classified breast tissue according to BIRADS categories. Bovis and Singh reached $71\%$ correctly classified mammograms, and Petroudi et al. achieved an overall correct classification of $76\%$ . Table  summarizes in more detail the results they obtained, including the results of our developed approach. It can be seen that Petroudi et al. obtained similar results to our MIAS database based evaluation, but with significant lesser results on BIRADS II and III and, hence, on the overall classification (column Total$_4$ ). Moreover, the table shows that Bovis has lower four class results on a smaller DDSM dataset, but higher overall low/high classification (column Total$_2$ ). Note, however, that a direct comparison is difficult because both have used different datasets. Bovis and Singh used $377$ DDSM MLO images (probably different from the ones used in our work), while Petroudi et al. used $132$ local (non-publicly available) CC/MLO images. Moreover, it is likely that the distribution over the various BIRADS categories is different in each experiment, and in turn, this could influence the results in the sense that a dataset with a distribution skewed towards BIRADS classes I and IV can be expected to show better results than a dataset with a distribution with a higher proportion of II and III category images. In our experiments a similar behaviour could be seen in the results obtained using the MIAS database and Expert A, who in comparison with Experts B and C used a high percentage of BIRADS I classifications.

Table 3.9: Comparison with existing work, with classification according to BIRADS categories. Numbers indicate the overall percentage of correct classification for each class, while the numbers in brackets indicate the percentage of images belonging to each BIRADS class with respect to the whole database. Total$_4$ indicates the overall percent of correct classification for the four classes problem, while Total$_2$ for the two classes problem, detailed in the columns Low and High.

---||--	B-I(%)	B-II (%)	B-III (%)	B-IV (%)	Total$_4$ (%)	Low (%)	High (%)	Total$_2$ (%)
---||--
---||-- Bovis [17]	(20)	(21)	(25)	(33)	71			97
---||-- Petroudi [137]	91	64	70	78	76	91	94
---||--
---||-- MIAS	91 (27)	84 (32)	89 (30)	73 (11)	86	89	94	91
---||-- DDSM	55 (13)	88 (40)	77 (31)	69 (16)	77	89	79	84