{"id":6500,"date":"2023-06-30T12:46:39","date_gmt":"2023-06-30T10:46:39","guid":{"rendered":"https:\/\/eurocc.nscc.sk\/?p=6500"},"modified":"2023-06-30T12:46:39","modified_gmt":"2023-06-30T10:46:39","slug":"meranie-strukturalnych-parametrov-kapsul-pouzitim-technik-umelej-inteligencie-ai-a-strojoveho-ucenia-ml","status":"publish","type":"post","link":"https:\/\/eurocc.nscc.sk\/en\/meranie-strukturalnych-parametrov-kapsul-pouzitim-technik-umelej-inteligencie-ai-a-strojoveho-ucenia-ml\/","title":{"rendered":"Measurement of microcapsule structural parameters using artificial intelligence (AI) and machine learning (ML)"},"content":{"rendered":"\n
\n
\n
\n
\n

Meranie \u0161truktur\u00e1lnych parametrov kaps\u00fal pou\u017eit\u00edm techn\u00edk umelej inteligencie (AI) a strojov\u00e9ho u\u010denia (ML)<\/strong><\/p>\n\n\n\n

<\/p>\n<\/div><\/div>\n\n\n\n

Cie\u013eom spolupr\u00e1ce medzi N\u00e1rodn\u00fdm kompeten\u010dn\u00fdm centrom pre HPC (NCC pre HPC) a \u00dastavom polym\u00e9rov (\u00daP) SAV bol n\u00e1vrh a implement\u00e1cia pilotn\u00e9ho softv\u00e9rov\u00e9ho rie\u0161enia pre automatick\u00e9 spracovanie obrazu frakci\u00ed polym\u00e9rnych mikrokaps\u00fal. Tieto mikrokapsuly sl\u00fa\u017eia ako obal pre pankreatick\u00e9 ostrov\u010deky tvoriace perspekt\u00edvne lie\u010divo na ochorenie diabetes mellitus (t.j. cukrovky 1. typu). Mikrokapsuly pozost\u00e1vaj\u00fa z pankreatick\u00fdch ostrov\u010dekov enkapsulovan\u00fdch do polopriepustnej polym\u00e9rnej membr\u00e1ny, ktor\u00e1 bola vyvinut\u00e1 na \u00dastave polym\u00e9rov SAV.<\/p>\n\n\n\n

<\/p>\n<\/div>\n\n\n\n

\n
\"\"<\/a>
Ilustra\u010dn\u00fd obr\u00e1zok<\/figcaption><\/figure>\n<\/div>\n<\/div>\n\n\n\n

<\/p>\n\n\n\n

Automatizovan\u00e9 rie\u0161enie je pre \u00daP SAV mimoriadne d\u00f4le\u017eit\u00e9 z h\u013eadiska \u010dasovej \u00faspory a zjednodu\u0161enia vyhodnocovania po\u010detn\u00fdch v\u00fdstupov z meran\u00ed, ako aj minimaliz\u00e1cie chyby, ktor\u00e1 sa m\u00f4\u017ee objavi\u0165 pri manu\u00e1lnom spracovan\u00ed. Obr\u00e1zky z optick\u00e9ho mikroskopu pri 4-n\u00e1sobnom zv\u00e4\u010d\u0161en\u00ed typicky obsahuj\u00fa jednu alebo nieko\u013eko mikrokaps\u00fal a s\u00fa vstupom do tr\u00e9ningu AI\/ML modelov. Obr\u00e1zky z optick\u00e9ho mikroskopu pri 2,5-n\u00e1sobnom zv\u00e4\u010d\u0161en\u00ed obsahuj\u00fa viac mikrokaps\u00fal, zvy\u010dajne tri a\u017e sedem. V takomto pr\u00edpade je nutn\u00e9 v prvom kroku identifikova\u0165 jednotliv\u00e9 mikrokapsuly. V procese inferencie sa z pr\u00edslu\u0161n\u00e9ho obr\u00e1zku vytvor\u00ed tzv. bin\u00e1rna maska, z ktorej sa n\u00e1sledne extrahuj\u00fa inform\u00e1cie o \u0161truktur\u00e1lnych parametroch, ak\u00fdmi s\u00fa predov\u0161etk\u00fdm vn\u00fatorn\u00fd a vonkaj\u0161\u00ed priemer kapsuly a hr\u00fabka jej membr\u00e1ny.<\/p>\n\n\n\n

Obrazov\u00fd materi\u00e1l je spracovan\u00fd v\u00a0dvoch krokoch. Prv\u00fdm krokom je lokaliz\u00e1cia nedefektn\u00fdch kaps\u00fal a\u00a0ich n\u00e1sledn\u00e9 vystrihnutie, druh\u00fdm je s\u00e9ria oper\u00e1ci\u00ed ved\u00facich k\u00a0samotn\u00e9mu ur\u010deniu \u0161truktur\u00e1lnych parametrov.<\/p>\n\n\n\n

DETEKCIA KAPS\u00daL<\/p>\n\n\n\n

Na detekciu kaps\u00fal bol pou\u017eit\u00fd model YOLOv5 [1] s predtr\u00e9novan\u00fdmi v\u00e1hami z datab\u00e1zy COCO128 [2]. Tr\u00e9ningov\u00e9 d\u00e1ta pozost\u00e1vali z 96 sn\u00edmok, ktor\u00e9 boli manu\u00e1lne anotovan\u00e9 pomocou n\u00e1stroja LabelImg [3]. Tr\u00e9ningov\u00e1 jednotka pozost\u00e1vala z 300 epoch, sn\u00edmky boli rozdelen\u00e9 do s\u00e1d po 16 a ich ve\u013ekos\u0165 bola nastaven\u00e1 na 640 pixelov. V\u00fdpo\u010dtov\u00fd \u010das jednej tr\u00e9ningovej jednotky na grafickej karte NVIDIA GeForce GTX 1650 bol pribli\u017ene 3.5 hodiny.<\/p>\n\n\n\n

Detekcia pomocou natr\u00e9novan\u00e9ho YOLOv5 modelu je prezentovan\u00e1 na Obr\u00e1zku 1. Spo\u013eahlivos\u0165 natr\u00e9novan\u00e9ho modelu, overen\u00e1 na 12 sn\u00edmkach, bola 96%, pri\u010dom priepustnos\u0165 na rovnakej grafickej karte bola pribli\u017ene 40 sn\u00edmok za sekundu.<\/p>\n\n\n

\n
\"\"<\/a>
Obr\u00e1zok 1: (a) vstupn\u00fd obraz z optickej mikroskopie; (b) detegovan\u00e1 kapsula; (c) v\u00fdrez detegovanej kapsuly pri 4-n\u00e1sobnom zv\u00e4\u010d\u0161en\u00ed; (d) vstupn\u00fd obraz z optickej mikroskopie; (e) detegovan\u00e1 kapsula; (f) v\u00fdrez detegovanej kapsuly pri 2,5-n\u00e1sobnom zv\u00e4\u010d\u0161en\u00ed.<\/figcaption><\/figure><\/div>\n\n\n

MERANIE \u0160TRUKTURALN\u00ddCH PARAMETROV KAPS\u00daL POMOCOU TECHN\u00cdK AI\/ML <\/p>\n\n\n\n

Bin\u00e1rne masky pre vn\u00fatorn\u00e9 a vonkaj\u0161ie \u010dasti kaps\u00fal boli z\u00edskan\u00e9 individu\u00e1lne, ako v\u00fdstup z hlbokej neur\u00f3novej siete architekt\u00fary U-Net [4], ktor\u00e1 bola vyvinut\u00e1 na spracovanie obrazu v biomedic\u00ednskych aplik\u00e1ci\u00e1ch. Na tr\u00e9ning pr\u00edslu\u0161n\u00fdch v\u00e1h bolo pou\u017eit\u00fdch 140 obr\u00e1zkov s kore\u0161ponduj\u00facimi maskami pre 4-n\u00e1sobn\u00e9 zv\u00e4\u010d\u0161enie optick\u00e9ho mikroskopu a rovnako 140 obr\u00e1zkov s kore\u0161ponduj\u00facimi maskami aj pre 2,5-n\u00e1sobn\u00e9 zv\u00e4\u010d\u0161enie. Tr\u00e9ningov\u00fd proces pozost\u00e1val z 200 epoch (ve\u013ekos\u0165 sady 16), pri\u010dom 10% z tr\u00e9ningov\u00fdch d\u00e1t bolo pou\u017eit\u00fdch na valid\u00e1ciu. Presnos\u0165 na testovacej sade, ktor\u00e1 pozost\u00e1vala zo 120 obr\u00e1zkov, presahovala 96%. Tr\u00e9ningov\u00fd proces trval 1,5 a\u017e 2 hodiny a na tr\u00e9novanie bol vyu\u017eit\u00fd HPC syst\u00e9m s uzlami typu IBM Power 7. Tento proces bolo nutn\u00e9 nieko\u013ekokr\u00e1t opakova\u0165. V\u00fdstupn\u00e9 bin\u00e1rne masky boli n\u00e1sledne postprocesovan\u00e9 oper\u00e1ciami \u201efill_holes\u201d [5] a \u201ewatershed\u201c [6], na z\u00edskanie \u010do najhlad\u0161\u00edch ov\u00e1lnych masiek. N\u00e1sledne bola na masky fitovan\u00e1 elipsa s vyu\u017eit\u00edm kni\u017enice \u201escikit-image measure\u201c [7], ktorej hlavn\u00e1 a ved\u013eaj\u0161ia os s\u00fa z\u00e1kladom pre samotn\u00fd v\u00fdpo\u010det \u0161truktur\u00e1lnych parametrov. Postupnos\u0165 t\u00fdchto krokov je prezentovan\u00e1 na Obr\u00e1zku 2.<\/p>\n\n\n

\n
\"\"<\/a>
Obr\u00e1zok 2: (a) vstupn\u00e1 obraz; (b) vn\u00fatorn\u00e1 bin\u00e1rna maska; (c) vonkaj\u0161ia bin\u00e1rna maska; (d) v\u00fdstupn\u00fd obraz s fitovan\u00fdmi elipsami elipsy na z\u00e1klade bin\u00e1rnych masiek.<\/figcaption><\/figure><\/div>\n\n\n

\u0160truktur\u00e1lne parametre z\u00edskan\u00e9 predikciou  AI\/ML modelu (zna\u010den\u00fd \u010falej ako \u201eU-Net\u201c) boli porovnan\u00e9 s hodnotami, ktor\u00fdmi disponovali pracovn\u00edci \u00daP SAV a boli z\u00edskan\u00e9 \u201emanu\u00e1lnym meran\u00edm\u201c v sn\u00edmkach. Ako \u010fal\u0161\u00ed, nez\u00e1visl\u00fd zdroj referen\u010dn\u00fdch d\u00e1t bol pou\u017eit\u00fd pr\u00edstup ozna\u010dovan\u00fd \u010falej ako \u201eRetinex\u201c, ktor\u00fd navrhol a implementoval RNDR. Andrej L\u00fa\u010dny, PhD. z Katedry aplikovanej informatiky Fakulty matematiky, fyziky a informatiky Univerzity Komensk\u00e9ho v Bratislave. Tento pr\u00edstup nevyu\u017e\u00edva neur\u00f3nov\u00e9 siete, ale vytv\u00e1ra kandid\u00e1tov na bin\u00e1rne masky agreg\u00e1ciou \u00fasekov s n\u00edzkou krivos\u0165ou, z\u00edskan\u00fdch pomocou \u201eretinex\u201c filtra [8] tak, aby vytvorili \u00fatvar s minim\u00e1lnou \u0161tandardnou odch\u00fdlkou krivosti. Tento pr\u00edstup je dobrou referenciou, nako\u013eko nevy\u017eaduje tr\u00e9novanie a je vysoko presn\u00fd, av\u0161ak na rozdiel od modelu U-net nie je plne automatizovan\u00fd, obzvl\u00e1\u0161\u0165 pre vn\u00fatorn\u00fa membr\u00e1nu kapsuly.<\/p>\n\n\n\n

Obr\u00e1zok 3 sumarizuje porovnanie v\u0161etk\u00fdch spomenut\u00fdch rie\u0161en\u00ed pre r\u00f4zne sady (\u201ebatches\u201c) z\u00a0experimentov pre obr\u00e1zky z\u00edskan\u00e9 pri 4-n\u00e1sobnom zv\u00e4\u010d\u0161en\u00ed.<\/p>\n\n\n

\n
\"\"<\/a>
(a)<\/figcaption><\/figure><\/div>\n\n
\n
\"\"<\/a>
(b)<\/figcaption><\/figure><\/div>\n\n
\n
\"\"<\/a>
(c)<\/figcaption><\/figure><\/div>\n\n\n

Obr\u00e1zok 3: <\/strong>(a) rozmer kaps\u00fal vypo\u010d\u00edtan\u00fd pre pr\u00edstupy U-net a Retinex ako priemer hlavnej a\u00a0ved\u013eaj\u0161ej osi elipsy (b) rozdiel medzi hlavnou osou elipsy pre vonkaj\u0161\u00ed priemer kaps\u00fal (c) rozdiel medzi ved\u013eaj\u0161ou osou elipsy pre vonkaj\u0161\u00ed priemer kaps\u00fal. V\u00a0obr\u00e1zkoch (b) a (c) je \u010dervenou \u010diarou zobrazen\u00e1 hrani\u010dn\u00e1 odch\u00fdlka akceptovate\u013en\u00e1 \u00daP SAV. Obr\u00e1zky z\u00a0optick\u00e9ho mikroskopu boli z\u00edskan\u00e9 pri 4-n\u00e1sobnom zv\u00e4\u010d\u0161en\u00ed.<\/p>\n\n\n\n

Zo z\u00edskan\u00fdch v\u00fdsledkov je mo\u017en\u00e9 skon\u0161tatova\u0165, \u017ee a\u017e na 4 sn\u00edmky (1.5%) v sade 194 pre ved\u013eaj\u0161iu os elipsy, s\u00fa v\u0161etky parametre v medziach akceptovate\u013enej presnosti, ktor\u00e1 bola definovan\u00e1 \u00daP SAV. Z Obr\u00e1zku 3(a) pozorujeme systematicky lep\u0161iu vz\u00e1jomn\u00fa zhodu medzi v\u00fdsledkami z\u00edskan\u00fdmi pr\u00edstupmi U-net a Retinex, \u010do m\u00f4\u017ee by\u0165 sp\u00f4soben\u00e9 bu\u010f aproxim\u00e1ciou tvaru kapsuly elipsou, ktor\u00e1 sa v \u201emanu\u00e1lnom vyhodnoten\u00ed\u201c \u00daP SAV neaplikovala, alebo in\u00fdm, systematick\u00fdm rozdielom pri vyhodnocovan\u00ed, a\/alebo chybou v \u201emanu\u00e1lnom vyhodnoten\u00ed\u201c. Kvalita U-net modelu m\u00f4\u017ee by\u0165 do bud\u00facna v\u00fdrazne zlep\u0161en\u00e1, hlavne roz\u0161\u00edren\u00edm tr\u00e9ningovej sady ako aj aplikovan\u00edm \u010fal\u0161ieho pre- a postprocesingu. Zhoda medzi \u201emanu\u00e1lnym vyhodnoten\u00edm\u201c a modelmi U-net \/ Retinex m\u00f4\u017ee by\u0165 zlep\u0161en\u00e1 harmoniz\u00e1ciou sp\u00f4sobu vyhodnocovania \u0161truktur\u00e1lnych parametrov kaps\u00fal z bin\u00e1rnych masiek.<\/p>\n\n\n\n

AI\/ML model bude nasaden\u00fd v predproduk\u010dnej f\u00e1ze ako cloudov\u00e9 rie\u0161enie na HPC syst\u00e9moch CS\u010c SAV. Inferencia a kontinu\u00e1lny tr\u00e9ning s prib\u00fadaj\u00facimi sn\u00edmkami nebude vy\u017eadova\u0165 invest\u00edciu do vysokov\u00fdkonn\u00fdch v\u00fdpo\u010dtov\u00fdch prostriedkov samotn\u00fdm \u00daP SAV. Produk\u010dn\u00e1 f\u00e1za, ktor\u00e1 presahuje r\u00e1mec pilotn\u00e9ho rie\u0161enia, uva\u017euje s integr\u00e1ciou tohto pr\u00edstupu do desktopovej aplik\u00e1cie.<\/p>\n\n\n\n

Zdroje<\/strong>:<\/p>\n\n\n\n

[1] https:\/\/github.com\/ultralytics\/yolov5<\/a><\/p>\n\n\n\n

[2] https:\/\/www.kaggle.com\/ultralytics\/coco128<\/p>\n\n\n\n

[3] https:\/\/github.com\/heartexlabs\/labelImg<\/a><\/p>\n\n\n\n

[4] https:\/\/lmb.informatik.uni-freiburg.de\/people\/ronneber\/u-net\/<\/a><\/p>\n\n\n\n

[5] https:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.ndimage.binary_fill_holes.html<\/a><\/p>\n\n\n\n

[6] https:\/\/scikit-image.org\/docs\/stable\/auto_examples\/segmentation\/plot_watershed.html<\/a><\/p>\n\n\n\n

[7] https:\/\/scikit-image.org\/docs\/stable\/api\/skimage.measure.html<\/a><\/p>\n\n\n\n

[8] D.J. Jobson, Z. Rahman, G.A. Woodell, IEEE Transactions on Image Processing 6 (7) 965-976, 1997.<\/p>\n\n\n\n


<\/p>\n\n\n\n

\n
Success-Stories<\/a><\/div>\n<\/div>\n\n\n
\"\"<\/a> Stop digit\u00e1lnym pochlebova\u010dom<\/a> 16 Aug<\/span> -<\/span> Vedky\u0148a pomocou superpo\u010d\u00edta\u010da prelomila zvyk modelov AI len pr\u00e1zdne chv\u00e1li\u0165 a nau\u010dila ich d\u00e1va\u0165 \u00faprimn\u00fa, kritick\u00fa sp\u00e4tn\u00fa v\u00e4zbu<\/span><\/div>
\"\"<\/a> Odvr\u00e1ten\u00e1 strana bezplatn\u00fdch hier<\/a> 11 Jul<\/span> -<\/span> Ako slovensk\u00fd t\u00edm s pomocou superpo\u010d\u00edta\u010da vyrie\u0161il r\u00e9bus extr\u00e9mne vz\u00e1cneho spr\u00e1vania hr\u00e1\u010dov<\/span><\/div>
\"\"<\/a> Ke\u010f AI klope na br\u00e1nu teol\u00f3gie<\/a> 11 Jun<\/span> -<\/span> Pre\u010do umel\u00e1 inteligencia patr\u00ed aj do teol\u00f3gie?<\/span><\/div><\/div>\n<\/div><\/div>\n","protected":false},"excerpt":{"rendered":"

The main aim of collaboration between the National Competence Centre for HPC (NCC HPC) and the Institute of Polymers of SAV (IP SAV) was design and implementation of a pilot software solution for automatic processing of polymer microcapsules images using artificial intelligence (AI) and machine learning (ML) approach. The microcapsules consist of semi-permeable polymeric membrane which was developed at the IP SAV.<\/p>","protected":false},"author":2,"featured_media":6513,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"templates\/template-full-width.php","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/posts\/6500"}],"collection":[{"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/comments?post=6500"}],"version-history":[{"count":7,"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/posts\/6500\/revisions"}],"predecessor-version":[{"id":6515,"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/posts\/6500\/revisions\/6515"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/media\/6513"}],"wp:attachment":[{"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/media?parent=6500"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/categories?post=6500"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/eurocc.nscc.sk\/en\/wp-json\/wp\/v2\/tags?post=6500"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}