3D Semi-Automatic Pollen Recognition Based on Palynological Knowledge

Alain Boucher1, Pablo J. Hidalgo2, Jordina Belmonte3,

Monique Thonnat1 and Carmen Galan2

1- INRIA, Sophia-Antipolis, France2- University of Córdoba (UCO), Spain3- Autonomous University of Barcelona (UAB), Cerdanyola del Vallès,Spain

9/08/2002 INRIA - UCO - UAB 2

Introduction

European project (1999 - 2001)Prevention and treatment of asthma and

allergyTwo aspects:

• IdentificationIdentification (types and concentrations) of the main aeroallergens (pollen grainspollen grains, dust)

•Forecast of the aeroallergen dispersion

Pollen recognition: two modulesImage acquisition of pollen grains in 3DPollen grain recognitionPollen grain recognition

9/08/2002 INRIA - UCO - UAB 3

Material and methods

Pollen grains are dyed with fuchsine fuchsine ((4

µg/100 ml)

Observation with a light microscopelight microscope (60x)

Automatic digitisation in 3D 3D

Database of more than 350 digitised grains 350 digitised grains

(30 different pollen types)

9/08/2002 INRIA - UCO - UAB 4

Main Pollen Types Studied and Similars

PoaceaeOlea ParietariaCupressaceae

Populus BrassicaceaeFraxinusLigustrumPhillyreaSalix

BroussonetiaMorusUrtica membranacea

CeltisCoriaria

9/08/2002 INRIA - UCO - UAB 5

3D pollen grain digitisation

3D acquisition of pollen grains set of images at different depths

Features may appear on different heights

• 100 optical sections• step = 0.5 microns

For each grain

9/08/2002 INRIA - UCO - UAB 6

Palynological knowledge

The system tries to mimic the palynologists Knowledge is necessary to identify pollen grains

9/08/2002 INRIA - UCO - UAB 7

Pollen recognition steps (1/2)

First step: coarse classification

Global measures on the grain (2D)Size, colour (RGB), shape, convexity, ...

Sampling date (external data for flowering season)

First estimations of possible types Sorted hypothesis list

9/08/2002 INRIA - UCO - UAB 8

Use of the pollinic calendar

MANRESA-PORATS. Mean weekly concentrations (P/m3) 1996-1998

0

5

10

15

20

25

30

35

40

45

50

55

60

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52

CHENOPODIACEAE-AMARANTHACEAE Plantago Celtis Coriaria POACEAE total

BELLATERRA-CUPRESSACEAE/Populus. Mean weekly concentrations (P/m3) 1994-1998

0

10

20

30

40

50

60

70

80

90

100

110

120

130

140

150

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52

CUPRESSÀCIES POPULUS

Recognition hypotheses includes the sampling dateMust take care of season variation

9/08/2002 INRIA - UCO - UAB 9

Pollen recognition steps (2/2)

Second step: fine classification

Search for specific characteristics (3D)

Need specific knowledge about pollen types

Driven by the hypothesis list test only the strongest hypotheses

Iterate and refine until no ambiguity remains

9/08/2002 INRIA - UCO - UAB 10

3D search in optical slices (key images - less blurred) 2D search in possible zones (regions of interest)

Search in a blurred image sequence

Image Mask

Interior Exine

Blur measure (SML) vs image number

9/08/2002 INRIA - UCO - UAB 11

Above central image Below central image

Sum of bright regions

Sum of dark regions

Example: Cupressaceae cytoplasm

Segmentation of bright regions Segmentation of dark regions

cytoplasm

9/08/2002 INRIA - UCO - UAB 12

Example: Olea reticulum

Network located on the external surface

Visible on top and bottom images Detection steps:

Check if the grain is reticulated Localise the reticulum (3D) Analyse the reticulum

9/08/2002 INRIA - UCO - UAB 13

Results

Test on a database of 350 pollen grains Reference images (pollen grains without dust and pollution) Simulation of the sampling date Leave-one-out method used for validation Results of recognition

Recognition rate between 4 allergenic types + others (5 classes) : 99,7%

Recognition rate between 31 pollen types + others (32 classes) : 77,7%

Test on a new set of images (different conditions) Low recognition rate between 4 allergenic pollen types (5 classes) : 45 % !! Problem of calibration and robustness for colour variation

Need to improve colour processing (more flexible system) Need to normalise image acquisition conditions

9/08/2002 INRIA - UCO - UAB 14

Future work: aerobiological images

Isolation of the pollen grains from dust