Focus-Science and Technology - - INNOVATIONS -

Google’s deep learn­ing AI has proven that it is more ac­cu­rate than pathol­o­gists at de­tect­ing breast can­cer that has spread to a pa­tient’s lymph nodes. The pres­ence or ab­sence of these ‘nodal metas­tases’ in­flu­ence a pa­tient’s prog­no­sis and treat­ment plan, so ac­cu­rate and fast de­tec­tion is im­por­tant. But in some cases, only 38 per cent of small metas­tases are picked up by pathol­o­gists when sam­ples are re­viewed un­der time con­straints, and right now, that pathol­o­gist’s ex­am­i­na­tion is the gold stan­dard in di­ag­no­sis of nodal metas­tases.

Google cus­tomised one of its ‘off-the-shelf’ deep learn­ing ap­proaches, call­ing it LYNA (LYmph Node As­sis­tant). Among other things, it was taught to ex­am­ine the images at dif­fer­ent mag­ni­fi­ca­tions, sim­i­lar to how a pathol­o­gist ex­am­ines slides. The al­go­rithm’s first test showed that LYNA was able to cor­rectly dis­tin­guish a slide with can­cer from a slide with­out 99 per cent of the time, even when the re­gions were too small to be de­tected by pathol­o­gists. In the sec­ond, six pathol­o­gists com­pleted a di­ag­nos­tic test with and with­out LYNA’s as­sis­tance. With LYNA’s help, the doc­tors found it ‘eas­ier’ to de­tect small metas­tases, and on aver­age the task took half as long. Pathol­o­gists work­ing with LYNA’s as­sis­tance were more ac­cu­rate than both unas­sisted pathol­o­gists and the LYNA al­go­rithm work­ing alone.

Google’s re­searchers sug­gest that al­go­rithms like LYNA could help with these iden­ti­fi­ca­tion tasks to al­low more time for pathol­o­gists to work on more com­plex di­ag­noses. But for now, fur­ther test­ing is needed to de­ter­mine if LYNA will work in real-life set­tings, which could in­volve a wider range of sam­ples from dif­fer­ent sites in the body.

The AI was shown a mi­cro­scope slide con­tain­ing lymph nodes (left), and it was able to cor­rectly iden­tify the tu­mor­ous re­gion by high­light­ing it in red (right)

Newspapers in English

Newspapers from UK

© PressReader. All rights reserved.