A novel ap­proach to early de­tec­tion of can­cer­ous lung nod­ules

BioSpectrum (India) - - BIO CONTENT - SINDHU RA­MACHAN­DRAN S., Prin­ci­pal Ar­chi­tect, Quest Global

Lung cancer is the lead­ing cause of cancer re­lated deaths in the world. As per the 2018 ‘State of Lung Cancer’ re­port by Amer­i­can Lung as­so­ci­a­tion, ev­ery two and a half min­utes some­one in the United States will be di­ag­nosed with lung cancer, with an es­ti­mated 234,030 new cases in the United States this year. The toll of lung cancer varies across na­tions. In In­dia too, it is the one of the lead­ing causes of cancer re­lated deaths.

A de­layed lung cancer di­ag­no­sis is as­so­ci­ated with a poor prog­no­sis. About 60 per cent of pa­tients dies within 1 year of di­ag­no­sis. Sur­gi­cal re­sec­tion of an early lung cancer has a favourable prog­no­sis. Af­ter re­sec­tion of a first stage bron­chogenic car­ci­noma, the 5-year sur­vival rate is 80 per cent to 90 per cent. There­fore, sig­nif­i­cant ef­forts tar­geted to­wards pre­ven­tion and early de­tec­tion are very cru­cial.

The chal­lenge of lung cancer de­tec­tion

Early de­tec­tion of lung nod­ules in pa­tients at risk for lung cancer can be achieved through Chest Com­puted To­mog­ra­phy (CT) screen­ing. This has shown to have im­proved sur­vival rate and such pro­grams have been im­ple­mented in clin­i­cal set­tings.

The ar­bi­trari­ness of shape, size and tex­ture of lung nod­ules is a chal­lenge when de­tect­ing the pres­ence of nod­ules. Ra­di­ol­o­gists an­a­lyze the chest CT scans with their ex­per­tise in hu­man anatomy and ex­pe­ri­ence in look­ing for struc­tures of in­ter­est. They iden­tify nod­ules in the image and an­no­tate them for fur­ther eval­u­a­tion. Rapid de­vel­op­ment of CT scan­ning tech­nolo­gies, in­creas­ing de­mand, more than hun­dreds of im­ages per exam, other pathol­ogy dis­tracts in the image are some of the chal­lenges faced by the ra­di­ol­o­gists in this process. More­over, repet­i­tive anal­y­sis and pre­cise de­tec­tion of nod­ules con­sumes a huge amount of ra­di­ol­o­gist’s ef­fort and time. Of­ten this leads to work pres­sure and heavy work­load re­sult­ing in many nod­ules be­ing missed in clin­i­cal prac­tice.

This has led to more fo­cus be­ing given to Com­puter Aided De­tec­tion (CAD) and di­ag­no­sis of lung nod­ules. CAD sys­tems can be very help­ful for achiev­ing a lung cancer screen­ing work­flow, which is ef­fi­cient as well as cost-ef­fec­tive.

A sin­gle CT scan of a pa­tient can con­tain 300 to 400 ax­ial im­ages. Look­ing for nod­ules in these scans re­quires a search through ap­prox­i­mately 300 ax­ial sections, each com­posed of over 260,000 pix­els. A 5-mm le­sion oc­cu­pies a small frac­tion of the image area and oc­curs within a back­ground of highly com­plex lung tis­sue. In con­ven­tional CAD sys­tems, the whole process of de­tec­tion in­volves gen­er­a­tion of can­di­date nod­ules and clas­si­fi­ca­tion of such nod­ules based on fea­tures which dif­fer­en­ti­ates a true nod­ule from a non-nod­ule. Since the fea­tures are hand­crafted, the ac­cu­racy and sen­si­tiv­ity of such sys­tems are low.

Us­ing deep learn­ing to achieve real time di­ag­no­sis

Of­ten de­scribed as a sub­set of ma­chine learn­ing, deep learn­ing deals with al­go­rithms in­spired by the struc­ture and func­tion of the hu­man brain, called ar­ti­fi­cial neu­ral net­works (ANN). It in­volves a com­pu­ta­tional model which is made to learn a prob­lem, by train­ing it on huge amounts of data as­so­ci­ated with that prob­lem. By get­ting trained on data, the model or net­work learns a set of pa­ram­e­ters or weights. The fully trained model can be de­ployed to make in­fer­ences us­ing the learned weights to do oper­a­tions like clas­si­fi­ca­tion, de­tec­tion etc. --in sim­ple terms the model mim­ics the brain. Deep learn­ing al­go­rithms called con­vo­lu­tional neu­ral net­works (CNN) are now used ex­ten­sively in solv­ing com­puter vi­sion prob­lems.

Con­clu­sion

Nod­ule de­tec­tion from CT scans us­ing deep learn­ing can give bet­ter per­for­mance, ac­cu­racy and sen­si­tiv­ity than other con­ven­tional de­tec­tion meth­ods. This in turn helps doc­tors to min­i­mize the time spent on the de­tailed vis­ual in­spec­tion of the CT im­ages. Their valu­able time can be used for com­mu­ni­cat­ing with pa­tients, plan­ning the treat­ment or for fur­ther re­search work. More­over, de­tect­ing the pres­ence of nod­ules early, leads to bet­ter cancer di­ag­no­sis and the sur­vival rate can be im­proved which makes the whole process re­ally re­ward­ing.

SINDHU RA­MACHAN­DRAN,

Prin­ci­pal Ar­chi­tect, QuEST Global

Newspapers in English

Newspapers from India

© PressReader. All rights reserved.