Deep Learning-Based Diagnostics Of Head And Neck Cancers Using Coherent Raman Scattering Microscopy

Bjarne Thorsted

Deep Learning-Based Diagnostics Of Head And Neck Cancers Using Coherent Raman Scattering Microscopy

Bjarne Thorsted

Institut for Biokemi og Molekylær Biologi

Publikation: Afhandling › Ph.d.-afhandling

Abstract

Hoved-halskræft (HHK) er en gruppe af relativt sjældne kræftformer, oftest
af typen pladecellekræft, som ikke viser typiske kræftlignende symptomer
i de tidlige stadier af sygdommen. De primære årsager årsager til HHK
er alkohol- og tobaksindtag, men også human papillomavirus (HPV) er
blevet en mere og mere almindelig årsag. Trods det, at behandlingen er
forholdsvis ukompliceret for tidlige stadier, så stiger dødeligheden markant
hvis patienten ikke søger lægehjælp før kræften har udviklet sig til de sene
stadier. Behandlingen består som regel af kirurgi og kemoradioterapi, hvor
kirurgien medfører risiko for varige mén i de sarte organer, der findes i
hoved-halsregionen. Grænsefladen mellem det raske væv og tumorvævet er
diskret og skifter gradvist, hvilket medfører at kirurgen ofte er afhængig
af intraoperative konsultationer (IOK) for at afgøre om nok væv er blevet
bortopereret. Denne process involverer at (noget af) vævet transporteres til
Afdeling for Klinisk Patologi hvor det undersøges makroskopisk, cryofryses,
snittes og vurderes med mikroskopi før et svar kan sendes retur til kirurgen
på operationsstuen. Ved at reducere denne ventetid vil man kunne nedbringe
udgifterne til operationen og potentielt øget patientflow samt en mindsket
arbejdsbyrde for Afdeling for Klinisk Patologi.
I denne afhandling blev anvendeligheden af et dybt neuralt netværk (DNN),
som et redskab til automatisk diagnosticering, undersøgt. Hypotesen er at
ikke-invasive billedmodaliteter, specifikt kohærent anti-Stokes Ramanspredning
(CARS)-, andenharmonisk generering (SHG)- og to-fotonseksitationsfluorescens-
mikroskopi, kan give fyldestgørende information om en labelfri
prøve til at et DNN kan lære at genkende forskellige vævstyper, hvis det
præsenteres for en database med sådanne billeder.
Studiet i dette projekt viser at kombinationen af CARS, SHG, og TPEF
leverer nok information til definitivt at kunne identificere adskillelige vigtige
detaljer ved both rask, dysplasisk og kræftramt væv. Derudover har studiet
også vist et et fuldt konvolutionelt neuralt netværk (FCNN) kan trænes
fra en relativt lille billeddatabase og opnå en imponerende nøjagtighed. Adskillige
hyperparametre blev evalueret og vi fandt at den bedste ydelse blev
opnået med en kombination af den særlige Swish-aktiveringsfunktion, batchnormalisering,
den regionsbaserede Tversky omkostningsfunktion samt
en række af inputdataforøgelsesalgoritmer, såsom tilfældig pixeldeletion
(dropout), rotation og elastisk deformering.
Resultaterne af dette studie viser at multi-modal imaging kombineret med
dyb lærings-baseret analyse i fremtiden kan hjælpe med at reducere patologens
arbejdsbyrde ved IOK’er samt at mere data er nødvendig for at øge ydeevnen
yderligere.

Originalsprog	Engelsk
Bevilgende institution	Syddansk Universitet
Vejledere/rådgivere	Brewer, Jonathan, Vejleder Godballe, Christian, Vejleder Larsen, Stine Rosenkilde, Vejleder, Ekstern person
Udgiver	Syddansk Universitet. Det Naturvidenskabelige Fakultet
Status	Udgivet - okt. 2020

Citationsformater

@misc{e0042ec9175d40b48e9e2d773e0a900f,

title = "Deep Learning-Based Diagnostics Of Head And Neck Cancers Using Coherent Raman Scattering Microscopy",

abstract = "head and neck cancers (HNCs) are rare cancers of mainly squamous cell carcinoma(SCC) that do not show symptoms typically associated with cancer inthe early stages of development. The main causes are alcohol and tobaccoconsumption with human papillomavirus (HPV) also becoming an increasinglycommon cause. While treatment is uncomplicated in the early stages,mortality significantly increases if patients do not seek medical attentionbefore the cancer has developed into the later stages. Treatment typicallyinvolves surgery and chemoradiotherapy, where the former carries the riskof damaging the delicate organs in the head and neck region. The interfacebetween healthy tissue and tumor is subtle and gradual, which means thatsurgeons are often dependent on intra-operative consultations (IOCs) duringoperation to determine whether enough tissue has been resected. This processinvolves transferring (a part of) the resected tissue to the Department ofPathology for macroscopic inspection, flash freezing, microsectioning andmicroscopic assessment before a response can be returned to the surgeonwaiting in the operating theater. Reducing this lag time would result incost reductions and potentially increased patient throughput as well as theworkload at the Department of Pathology.In this thesis the applicability of a deep neural network (DNN) as anautomated diagnosing tool was investigated. The hypothesis being that noninvasiveimaging modalities, specifically coherent anti-Stokes Raman scattering(CARS), second-harmonic generation (SHG), and two-photon excited fluorescence(TPEF) microscopy, could provide adequate information about alabel-free sample for a DNN to learn to recognize different tissue types ifpresented with a database of such images.The study in this project shows that the combination of CARS, SHG, andTPEF deliver enough information to positively identify several importantfeatures of both healthy, dysplasic, and cancerous tissue. Furthermore, italso showed that a fully convolutional neural network (FCNN) can be trainedfrom a relatively small image database and achieve an impressive accuracy.Several hyper parameters were evaluated and we found that the highestperformance was achieved with a combination of the novel Swish activationfunction, batch normalization, the region-based Tversky loss function and ahost of input data augmentation algorithms including, but not limited to,random pixel deletion (dropout), rotation and elastic deformation.The results of this study shows that multi-modal imaging in combinationwith deep learning-based analysis in the future can help reduce the pathologistsworkload during IOCs and that more data is needed to improve theperformance.",

author = "Bjarne Thorsted",

year = "2020",

month = oct,

language = "English",

publisher = "Syddansk Universitet. Det Naturvidenskabelige Fakultet",

address = "Denmark",

school = "SDU",

}

TY - GEN

T1 - Deep Learning-Based Diagnostics Of Head And Neck Cancers Using Coherent Raman Scattering Microscopy

AU - Thorsted, Bjarne

PY - 2020/10

Y1 - 2020/10

N2 - head and neck cancers (HNCs) are rare cancers of mainly squamous cell carcinoma(SCC) that do not show symptoms typically associated with cancer inthe early stages of development. The main causes are alcohol and tobaccoconsumption with human papillomavirus (HPV) also becoming an increasinglycommon cause. While treatment is uncomplicated in the early stages,mortality significantly increases if patients do not seek medical attentionbefore the cancer has developed into the later stages. Treatment typicallyinvolves surgery and chemoradiotherapy, where the former carries the riskof damaging the delicate organs in the head and neck region. The interfacebetween healthy tissue and tumor is subtle and gradual, which means thatsurgeons are often dependent on intra-operative consultations (IOCs) duringoperation to determine whether enough tissue has been resected. This processinvolves transferring (a part of) the resected tissue to the Department ofPathology for macroscopic inspection, flash freezing, microsectioning andmicroscopic assessment before a response can be returned to the surgeonwaiting in the operating theater. Reducing this lag time would result incost reductions and potentially increased patient throughput as well as theworkload at the Department of Pathology.In this thesis the applicability of a deep neural network (DNN) as anautomated diagnosing tool was investigated. The hypothesis being that noninvasiveimaging modalities, specifically coherent anti-Stokes Raman scattering(CARS), second-harmonic generation (SHG), and two-photon excited fluorescence(TPEF) microscopy, could provide adequate information about alabel-free sample for a DNN to learn to recognize different tissue types ifpresented with a database of such images.The study in this project shows that the combination of CARS, SHG, andTPEF deliver enough information to positively identify several importantfeatures of both healthy, dysplasic, and cancerous tissue. Furthermore, italso showed that a fully convolutional neural network (FCNN) can be trainedfrom a relatively small image database and achieve an impressive accuracy.Several hyper parameters were evaluated and we found that the highestperformance was achieved with a combination of the novel Swish activationfunction, batch normalization, the region-based Tversky loss function and ahost of input data augmentation algorithms including, but not limited to,random pixel deletion (dropout), rotation and elastic deformation.The results of this study shows that multi-modal imaging in combinationwith deep learning-based analysis in the future can help reduce the pathologistsworkload during IOCs and that more data is needed to improve theperformance.

AB - head and neck cancers (HNCs) are rare cancers of mainly squamous cell carcinoma(SCC) that do not show symptoms typically associated with cancer inthe early stages of development. The main causes are alcohol and tobaccoconsumption with human papillomavirus (HPV) also becoming an increasinglycommon cause. While treatment is uncomplicated in the early stages,mortality significantly increases if patients do not seek medical attentionbefore the cancer has developed into the later stages. Treatment typicallyinvolves surgery and chemoradiotherapy, where the former carries the riskof damaging the delicate organs in the head and neck region. The interfacebetween healthy tissue and tumor is subtle and gradual, which means thatsurgeons are often dependent on intra-operative consultations (IOCs) duringoperation to determine whether enough tissue has been resected. This processinvolves transferring (a part of) the resected tissue to the Department ofPathology for macroscopic inspection, flash freezing, microsectioning andmicroscopic assessment before a response can be returned to the surgeonwaiting in the operating theater. Reducing this lag time would result incost reductions and potentially increased patient throughput as well as theworkload at the Department of Pathology.In this thesis the applicability of a deep neural network (DNN) as anautomated diagnosing tool was investigated. The hypothesis being that noninvasiveimaging modalities, specifically coherent anti-Stokes Raman scattering(CARS), second-harmonic generation (SHG), and two-photon excited fluorescence(TPEF) microscopy, could provide adequate information about alabel-free sample for a DNN to learn to recognize different tissue types ifpresented with a database of such images.The study in this project shows that the combination of CARS, SHG, andTPEF deliver enough information to positively identify several importantfeatures of both healthy, dysplasic, and cancerous tissue. Furthermore, italso showed that a fully convolutional neural network (FCNN) can be trainedfrom a relatively small image database and achieve an impressive accuracy.Several hyper parameters were evaluated and we found that the highestperformance was achieved with a combination of the novel Swish activationfunction, batch normalization, the region-based Tversky loss function and ahost of input data augmentation algorithms including, but not limited to,random pixel deletion (dropout), rotation and elastic deformation.The results of this study shows that multi-modal imaging in combinationwith deep learning-based analysis in the future can help reduce the pathologistsworkload during IOCs and that more data is needed to improve theperformance.

M3 - Ph.D. thesis

PB - Syddansk Universitet. Det Naturvidenskabelige Fakultet

ER -

Deep Learning-Based Diagnostics Of Head And Neck Cancers Using Coherent Raman Scattering Microscopy

Abstract

Fingeraftryk

Citationsformater