{"id":10528,"date":"2020-01-22T11:07:47","date_gmt":"2020-01-22T08:07:47","guid":{"rendered":"https:\/\/umram.bilkent.edu.tr\/?p=10528"},"modified":"2020-01-22T11:44:41","modified_gmt":"2020-01-22T08:44:41","slug":"transfer-learning-approach","status":"publish","type":"post","link":"https:\/\/umram.bilkent.edu.tr\/index.php\/2020\/01\/22\/transfer-learning-approach\/","title":{"rendered":"Transfer\u2010Learning Approach for Accelerated MRI Using Deep Neural Networks"},"content":{"rendered":"

Salman Ul Hassan Dar<\/span><\/a>,\u00a0<\/span>Muzaffer \u00d6zbey<\/span><\/a>,\u00a0<\/span>Ahmet Burak \u00c7atl\u0131<\/span><\/a>,\u00a0and\u00a0<\/span>Tolga \u00c7ukur<\/span><\/a>\u00a0published\u00a0an\u00a0article\u00a0entitled\u00a0“<\/span>A Transfer\u2010Learning Approach for Accelerated MRI Using Deep Neural Networks<\/span><\/a>\" in\u00a0Magnetic\u00a0Resonance\u00a0in\u00a0Medicine\u00a0at UMRAM.<\/span>\u00a0<\/span><\/p>\n

Abstract<\/span>\u00a0<\/span><\/p>\n

Purpose<\/span>\u00a0<\/span><\/p>\n

Neural\u00a0networks\u00a0have\u00a0received\u00a0recent\u00a0interest\u00a0for\u00a0reconstruction\u00a0of\u00a0undersampled\u00a0MR\u00a0acquisitions.\u00a0Ideally, network\u00a0performance\u00a0should\u00a0be\u00a0optimized\u00a0by\u00a0drawing\u00a0the\u00a0training\u00a0and\u00a0testing\u00a0data\u00a0from\u00a0the\u00a0same\u00a0domain.\u00a0In\u00a0practice,\u00a0however,\u00a0large\u00a0datasets\u00a0comprising\u00a0hundreds\u00a0of\u00a0subjects\u00a0scanned\u00a0under\u00a0a\u00a0common\u00a0protocol\u00a0are\u00a0rare.\u00a0The\u00a0goal\u00a0of\u00a0this\u00a0study\u00a0is\u00a0to\u00a0introduce\u00a0a transfer\u2010learning\u00a0approach\u00a0to\u00a0address\u00a0the\u00a0problem of data\u00a0scarcity\u00a0in\u00a0training\u00a0deep\u00a0networks\u00a0for\u00a0accelerated\u00a0MRI.<\/span>\u00a0<\/span><\/p>\n

Methods<\/span>\u00a0<\/span><\/p>\n

Neural\u00a0networks\u00a0were\u00a0trained\u00a0on\u00a0thousands\u00a0(upto\u00a04\u00a0thousand) of\u00a0samples\u00a0from\u00a0public\u00a0datasets\u00a0of\u00a0either\u00a0natural\u00a0images\u00a0or\u00a0brain\u00a0MR\u00a0images.\u00a0The\u00a0networks\u00a0were\u00a0then\u00a0fine\u2010tuned\u00a0using\u00a0only\u00a0tens\u00a0of\u00a0brain\u00a0MR\u00a0images\u00a0in a\u00a0distinct\u00a0testing\u00a0domain. Domain\u2010transferred\u00a0networks\u00a0were\u00a0compared\u00a0to\u00a0networks\u00a0trained\u00a0directly\u00a0in\u00a0the\u00a0testing\u00a0domain. Network\u00a0performance\u00a0was\u00a0evaluated\u00a0for\u00a0varying\u00a0acceleration\u00a0factors\u00a0(4\u201010),\u00a0number\u00a0of\u00a0training\u00a0samples\u00a0(0.5\u20104k),\u00a0and\u00a0number\u00a0of\u00a0fine\u2010tuning\u00a0samples\u00a0(0\u2010100).<\/span>\u00a0<\/span><\/p>\n

Results<\/span>\u00a0<\/span><\/p>\n

The\u00a0proposed\u00a0approach\u00a0achieves\u00a0successful\u00a0domain transfer\u00a0between\u00a0MR\u00a0images\u00a0acquired\u00a0with\u00a0different\u00a0contrasts\u00a0(T<\/span>1<\/span>\u2010\u00a0and\u00a0T<\/span>2<\/span>\u2010weighted\u00a0images)\u00a0and\u00a0between\u00a0natural\u00a0and\u00a0MR\u00a0images\u00a0(ImageNet\u00a0and\u00a0T<\/span>1<\/span>\u2010\u00a0or\u00a0T<\/span>2<\/span>\u2010weighted\u00a0images). Networks\u00a0obtained\u00a0via\u00a0transfer\u00a0learning\u00a0using\u00a0only\u00a0tens\u00a0of\u00a0images\u00a0in\u00a0the\u00a0testing\u00a0domain\u00a0achieve\u00a0nearly\u00a0identical\u00a0performance\u00a0to\u00a0networks\u00a0trained\u00a0directly\u00a0in\u00a0the\u00a0testing\u00a0domain\u00a0using\u00a0thousands\u00a0(upto\u00a04\u00a0thousand) of\u00a0images.<\/span>\u00a0<\/span><\/p>\n

Conclusion<\/span>\u00a0<\/span><\/p>\n

The<\/span>\u00a0<\/span>proposed<\/span>\u00a0<\/span>approach<\/span>\u00a0<\/span>might<\/span>\u00a0<\/span>facilitate<\/span>\u00a0<\/span>the<\/span>\u00a0<\/span>use<\/span>\u00a0of\u00a0<\/span>neural<\/span>\u00a0<\/span>networks<\/span>\u00a0<\/span>for<\/span>\u00a0MRI\u00a0<\/span>reconstruction<\/span>\u00a0<\/span>without<\/span>\u00a0<\/span>the<\/span>\u00a0<\/span>need<\/span>\u00a0<\/span>for<\/span>\u00a0<\/span>collection<\/span>\u00a0of\u00a0<\/span>extensive<\/span>\u00a0<\/span>imaging<\/span>\u00a0<\/span>datasets<\/span>.<\/span>\u00a0<\/span><\/p>\n<\/div><\/div><\/div><\/div><\/div><\/section>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":10120,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[256],"tags":[],"_links":{"self":[{"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/posts\/10528"}],"collection":[{"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/comments?post=10528"}],"version-history":[{"count":10,"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/posts\/10528\/revisions"}],"predecessor-version":[{"id":10666,"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/posts\/10528\/revisions\/10666"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/media\/10120"}],"wp:attachment":[{"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/media?parent=10528"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/categories?post=10528"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/umram.bilkent.edu.tr\/index.php\/wp-json\/wp\/v2\/tags?post=10528"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}